Breast Imaging

We performed a preliminary investigation to determine how similar radiologists’ interpretations of screening mammograms are. Our dataset consisted of 50 cancer cases and 50 normal cases that were read by 50 radiologists. We computed sensitivity, specificity, and interpretation on a case-by-case basis to study similarity between pairs of radiologists. We failed to find any pairs of radiologists who read all the cases, only the normal cases, or only the cancer cases the same. There were very few radiologists who read both cancer cases and normal cases in a similar manner. Even radiologists who had similar sensitivities or similar specificities differed substantially on the interpretation of individual cases. Our data indicate that there may be an underlying variability between radiologists in terms of image features used to detect cancers and when a false detection is made. This underlying variability may make the development and implementation of model observers more difficult.

The Tomosynthesis Mammography Imaging Screening Trial (TMIST) Lead-In Study is a randomized screening trial that aims to compare the performance of standard two-dimensional full-field digital mammography (FFDM) and tomosynthesis at five sites in Canada, with multiple vendors’ platforms and a target enrollment of 6300 women. To characterize and monitor the image quality of the tomosynthesis systems in the trial, a quality control (QC) program has been developed, including semi-annual physics tests, and daily tests performed on a phantom imaged by the radiographer.Here we describe the test regimen and phantoms and present initial results. The physics tests include measurement of image quality parameters in the reconstructed tomographic slices and evaluation of the AEC performance by measuring signal difference to noise ratio (SDNR) of a low contrast simulated lesion. The physics tests have been performed on a GE Senoclaire, two Hologic Selenia Dimensions and two Siemens Mammomat Inspiration units. In addition to the physics tests, we have remotely collected 15 months of daily QC data on the GE unit and 2 and 3 months on the Hologic units.

Digital breast tomosynthesis (DBT) screening outcomes are sustainable over consecutive years with significant reductions in recall and increasing cancers per recalled patients compared to screening with digital mammography alone (DM). There is a prevalence effect with a reduction in cancer detection at the second round of screening that is no longer present at the third round. There is a non-statistically significant trend of decreased interval cancers with DBT compared to DM alone screening. Early data on the implementation of synthetic 2D (s2D) imaging coupled with DBT shows maintenance of screening outcomes with reduction in radiation dose compared to DM/DBT screening .

Quantum noise is a signal-dependent, Poisson-distributed noise and the dominant noise source in digital mammography. Quantum noise removal or equalization has been shown to be an important step in the automatic detection of microcalcifications. However, it is often limited by the difficulty of robustly estimating the noise parameters on the images. In this study, a nonparametric image intensity transformation method that equalizes quantum noise in digital mammograms is described. A simple Look-Up-Table for Quantum Noise Equalization (LUT-QNE) is determined based on the assumption that noise properties do not vary significantly across the images. This method was evaluated on a dataset of 252 raw digital mammograms by comparing noise statistics before and after applying LUT-QNE. Performance was also tested as a preprocessing step in two microcalcification detection schemes. Results show that the proposed method statistically significantly improves microcalcification detection performance.

Convolutional neural networks (CNNs) have shown to be powerful for classification of image data and are increasingly used in medical image analysis. Therefore, CNNs might be very suitable to detect microcalcifications in mammograms. In this study, we have configured a deep learning approach to fulfill this task. To overcome the large class imbalance between pixels belonging to microcalcifications and other breast tissue, we applied a hard negative mining strategy where two CNNs are used. The deep learning approach was compared to a current state-of-the-art method for the detection of microcalcifications: the cascade classifier. Both methods were trained on a large training set including 11,711 positive and 27 million negative samples. For testing, an independent test set was configured containing 5,298 positive and 18 million negative samples. The mammograms included in this study were acquired on mammography systems from three manufactures: Hologic, GE, and Siemens. Receiver operating characteristics analysis was carried out. Over the whole specificity range, the CNN approach yielded a higher sensitivity compared to the cascade classifier. Significantly higher mean sensitivities were obtained with the CNN on the mammograms of each individual manufacturer compared to the cascade classifier in the specificity range of 0 to 0.1. To our knowledge, this was the first study to use a deep learning strategy for the detection of microcalcifications in mammograms.

Presentation of images similar to a new unknown lesion can be helpful in medical image diagnosis and treatment planning. We have been investigating a method to retrieve relevant images as a diagnostic reference for breast masses on mammograms and ultrasound images. For retrieval of visually similar images, subjective similarities for pairs of masses were determined by experienced radiologists, and objective similarity measures were computed by modeling the subjective similarity space using multidimensional scaling (MDS). In this study, we investigated the similarity measure for masses on breast ultrasound images based on MDS and an artificial neural network and examined its usefulness in image retrieval. For 666 pairs of masses, correlation coefficient between the average subjective similarities and the MDS-based similarity measure was 0.724. When one to five images were retrieved, average precision in selecting relevant images, i.e., pathology-matched images for benign/malignant index image, was 0.778, indicating the potential utility of the proposed MDS-based similarity measure.

In this paper, we employ a deep Convolutional Neural Network (CNN) for the classification of regions of interest of malignant soft tissue lesions in mammography and show that it performs on par to experienced radiologists. The CNN was applied to 398 regions of 5$$\,\times \,$$5 cm, half of which contained a malignant lesion and the other half depicted suspicious regions in normal mammograms detected by a traditional CAD system. Four radiologists participated in the study. ROC analysis was used for evaluating results. The AUC of CNN was 0.87, which was higher than the mean AUC of the radiologists (0.84), though the difference was not significant.

We evaluated the diagnostic performance of a novel image processing technique of synthetic MMG (SMMG) as 2D-like visualization approach with and without DBT slice images (DBT) compared with that of 2D MMG (MMG) alone. With one-view MMG and DBT, the radiation doses, utilizing the ACR phantom 156, were 1.20 mGy and 1.80 mGy. The number of the cases was 108. MMG, SMMG, and DBT slice images were evaluated independently by 4 readers utilizing ROC analysis and diagnostic performance. SMMG plus DBT demonstrated higher area under the curve (AUC) and superior diagnostic accuracy with sensitivity, specificity, and NPV compared with SMMG and MMG alone (p < 0.05). In addition, a 40 % decrease of radiation dose with SMMG plus DBT compared with MMG plus DBT as the current setting will enable us to apply a two-view SMMG plus DBT in breast cancer screening instead of MMG.

In Japan, soft-copy diagnoses shift rapidly, and the method for facilities applying it is necessary for soft-copy diagnosis. Digital mammography has high resolution and a large matrix size. On a monitor, the image is either displayed partially at 1:1 pixel mapping or narrowed to fit the screen, resulting in loss of image quality. Therefore, we developed a digital phantom for soft-copy diagnosis in digital mammography. This phantom is like the Contrast Detail Phantom and comprises 12 different shapes and eight different brightness levels. It becomes one group of nine, and each signal is located at the prime number coordinate from a central signal coordinate. Visual evaluation refers to the visibility of the nine signal coordinates when the image is adjusted to fit the monitor’s display. Although digital phantoms have been implemented at 120 facilities, at two other facilities, viewer problems were detected.

The EU funded project EUtempe-RX to develop 12 modules for training of medical physics experts (MPEs) in diagnostic and interventional radiology. Each course module provided 80 h of blended learning (a mixture of online and face-to-face training). The effectiveness of high-level blended learning for training MPEs up to EQF level 8 was tested on optimisation in mammography and general radiology. The training methods were evaluated using a questionnaire (89 % response rate) and reviewing participants’ proposed optimisation studies. The online training was the most highly rated part of the module. The participants produced a wide range of feasible and interesting optimisation proposals. All questionnaire responders intend to undertake their study. Overall, the proposals showed a good understanding of the process of optimisation, but some showed weaknesses in applying the results clinically. The blended learning approach showed potential for training MPEs to undertake successful optimisation projects.

We propose a simplified method for performing an FROC observer study by using regions of interest (ROIs) of CDMAM phantom images as case sample images. This new method would help researchers to plan and perform FROC observer studies in shorter time, with smaller errors and reduced complexity. As examples of experimental procedure, the digital images obtained by a direct-conversion flat-panel detector (FPD) system in three imaging conditions with different levels of patient dose were used to compare the diagnostic accuracies for low-contrast signal detection. There was a high correlation between the average figure of merit (FOM) obtained by our proposed method and inverse image quality figure (inv. IQF) calculated by CDMAM Analyser 1.55 in each dose level (r = 0.98).

The aim is to compare equivocal breast findings in digital breast tomosynthesis combined with digital mammography (DM+DBT) compared with DM alone for readers with different levels of DBT experience. Fifty cases (23 normal/benign, 27 cancer) were rated by 26 experienced breast radiologists (9, 9 and 8 individuals had no-, workshop- and clinical-DBT experience respectively) using a 5-point scoring scale (1 = normal, 5 = malignant). Ratings were compared between DM and DM+DBT for normal and cancer cases for all readers and for each sub-group of readers. Equivocal findings were 35 % of the total cancer-cases for DM compared with 24 % for DM+DBT for all readers grouped together with reductions also seen for each of the sub-groups. 31 % of normal cases were scored as equivocal in DM compared with 19 % in DM+DBT. Addition of DBT to DM reduces the number of equivocal breast findings regardless of the level of previous DBT experience.

We propose a method for estimating mammary gland composition and report the examined results to find the better conditions to improve the precision of the estimation. We use a cadmium telluride series (CdTe-series) detector as a photon-counting mammography detector in this study, since CdTe-series detectors detect photons in a wide energy range and provide highly accurate energy discrimination. An imaging system using a CdTe-series detector is simulated by MATLAB. We divide the spectrum of an X-ray, which is transmitted a phantom, into three energy bins and calculate the corresponding linear attenuation coefficients from the numbers of input and output photons. These linear attenuation coefficients are plotted in a three-dimensional (3D) scatter plot. Using this 3D scatter plot, we estimate the mammary gland composition and determine the optimal conditions to estimate.

Historically, with film/screen mammography systems image quality as a function of breast thickness was constrained by the necessity to maintain constant dose to the film across all thicknesses. With digital units, no such constraint is placed upon the AEC system and different manufacturers have designed their own breast thickness compensation schemes. In this study, threshold contrast detail detectability measurements were made at three different simulated breast thicknesses using a CDMAM phantom on analogue and digital mammography units. The purpose was to understand how the design of the different AEC systems affected the image quality for thicker breasts in particular, when compared with that for a standard breast thickness on the same unit. The results showed that relative image quality for thicker breasts compared with standard breasts varies greatly with unit type, depending on the thickness compensation scheme implemented by the particular AEC system. The work highlights an urgent need for a more evidence based approach to determining optimal task-based image quality as a function of breast thickness.

Higher recall rates have been related to increased false positive decisions, causing significant psychological and economical costs for both screened women and the mammography screening service respectively. This study compares breast readers’ performance in a laboratory setting under varying levels of recall rates. Four experienced radiologists volunteered to read a single test set of 200 mammographic cases over three separate conditions. The test set contained of 180 normal and 20 abnormal cases and the participants were asked to identify each case that required to be recalled in line with three different target recall rates: control (unspecified or free recall (first read)), 15 % (second read) and 10 % (third read). Readers were required to mark the location of any malignancies using custom made detection software. The recall rates for the control condition ranged between 18.5 % and 34 %. Statistically significant differences were observed in sensitivity for control (median = 0.85) vs 15 % (median = 0.65, z = -2.381, P = 0.017), 15 % vs 10 % (median = 0.55, z = -2.428, P = 0.015) and control vs 10 % (z = -2.381, P = 0.017). ROC AUC was significantly different for control (median = 0.84) vs 15 % (median = 0.79, z = -2.381, P = 0.017) and 15 % vs 10 % (median = 0.75, z = -2.381, P = 0.017). Specificity significantly improved at lower recall rate of 10 % (median = 0.95) vs 15 % (median = 0.92, z = -2.428, P = 0.017). Setting specific target recall rates for readers significantly limited their performance in correctly identifying cancers. In this study, decreasing the number of recalled cases down to 10 %, significantly reduced cancer detection, with a significant improvement in specificity (P ≤ 0.05).

The Voxel Imaging PET (VIP) Pathfinder project presents a new approach for the design of nuclear medicine imaging diagnostic devices by using highly segmented pixel CdTe sensors. State-of-the-art PET devices are made from scintillator crystals which have an energy resolution of about 10 %. Because of their limited energy resolution, the scatter fraction is relatively large. Their limited spatial resolution will introduce a parallax error and deficiency in the depth of interaction (DOI) determination (> 5 mm FWHM) and therefore a wrong line of response (LOR). In this study, we present a design for a breast dedicated PET based on CdTe detectors which have an energy resolution of about 1 % and can be easily segmented into small voxels for optimal spatial resolution to provide a solution for these setbacks. This simulation study will assess the advantages of this design for breast imaging purposes.

To examine the utility of employing breast screening test sets internationally the data of 1,009 radiologists from the USA, UK and other European countries were examined as they inspected 20 carefully selected difficult recent screening cases. Some 720 UK radiologists, 247 American and 42 European radiologists took part. Whilst similar sensitivity scores between the three groups were found, the main difference was the lower specificity of the American radiologists reflecting their different recall clinical practice. It is argued that using test sets internationally provides participants with useful comparative performance information whilst also providing data on how the same cases are interpreted by radiologists from different countries.

Contrast-enhanced digital breast tomosynthesis (CEDBT) may improve contrast-enhanced lesion conspicuity and relative contrast quantification by improving three-dimensional visualization of lesion morphology, and reducing the integration of attenuation information along the axial direction. Improved visualization of patterns of contrast-enhancement and improved iodine quantification may help differentiate between malignant and benign enhancing lesions. The dependence of dual-energy contrast-enhanced lesion detectability on imaging chain design is investigated. Lesion detectability and relative iodine quantification is comparable for subtraction in either reconstruction or projection domains for both phantom and patient images. SART generally produces greater SDNR than FBP, and scatter correcting projections further improves SDNR.

The aim of this work is the evaluation of the software BreastSimulator, as a tool for the creation of 3D uncompressed breast digital models and for the simulation and the optimization of Computed Tomography (CT) equipment. Three 3D digital breast phantoms were created, having different sizes and with realistic anatomical features. We calculated 2D X-ray CT projections simulating a breast tomogram with a dedicated cone-beam CT scanner. From the reconstructed CT slices, the power-law exponent, has been evaluated from the Noise Power Spectrum function S(f) = α/fβ. The results were then verified by comparison against clinical CT and published data. The preliminary results of this study showed that the simulated model complexity may reproduce the real anatomical complexity of the breast tissues as described, in terms of β values, since the measured β coefficients are close to that of clinical CT data from a dedicated breast CT scanner.

The aim of this study was to investigate the effect of dose on the detection of micro-calcification clusters in breast images using planar mammography and digital breast tomosynthesis (DBT). Planar and DBT images were created from mathematical models of breasts with and without inserted clusters of 5 identical calcifications. Regions of interest from the images were used in a series of 4-alternative forced choice human observer experiments using the clusters as targets. Three calcification diameters were used for each imaging condition. The threshold diameter required for micro-calcification detection was determined for a detection rate of 92.5 % at mean glandular doses of 1.25, 2.5, and 5 mGy. The measured threshold micro-calcification diameter was lower for planar mammography than for the DBT modality. The threshold micro-calcification diameter decreased with increasing dose for planar and DBT imaging. The image modality used had a larger effect on the threshold diameter than the dose change considered.

The Monte Carlo code PENELOPE (version 2014) and its software package PenEasy have been employed to closely simulate a commercial mammography system: Selenia Dimensions system (Hologic, Bedford, Mass) for dosimetric study. Normalized glandular dose (D_gN) was derived for specific filter combinations including tungsten/aluminum (W/Al), tungsten/rhodium (W/Rh) and tungsten/silver (W/Ag) for a range of clinically used voltages (kV). Breast models with thickness 2 cm, 4.5 cm and 8 cm and respective glandularity 100 %, 50 % and 0.1 % were included in the simulations. Derived D_gN values were in good agreement with well-accepted published values, with mean differences of 1.30 %, 2.17 % and 0.17 % for the W/Al, W/Rh and W/Ag anode/filter combinations, respectively. In the next phase of the study, this validated model will be modified to perform patient-specific dosimetry with an improved breast model and more realistic simulation of commercial mammography systems.

Mammography is one of the most effective techniques for early detection of breast cancer. The quality of the image may suffer from poor resolution or low contrast, which can effect the efficiency of radiologists. In order to improve the visual quality of mammograms, this paper introduces a new mammographic image enhancement algorithm. Firstly an intensity based nonlinear transformation is used for reducing the background tissue intensity, and secondly adaptive local contrast enhancement is realized based on local standard deviation and luminance information. The proposed method can obtain improved performance compared to alternative methods both covering objective and subjective aspects, based on 45 images. Experimental results demonstrate that the proposed algorithm can improve the contrast effectively and enhance lesion information (microcalcifications and/or masses).

Architectural distortion in mammograms is the most frequently missed finding among breast cancer findings, the improvement of detection accuracy in existing commercial CAD software remains a challenge. In this study, in order to improve the detection accuracy of architectural distortion in mammography, we propose a hybrid automatic detection method that combines with the enhancement method of the concentration of line structure and massive pattern. In the method, the detection of the concentration of the line structure is conducted by the adaptive Gabor filter, and the enhancement of the massive pattern is performed by the iris filter. The concentration index is calculated from these filtered images; the lesion candidate regions are obtained. As for false positive (FP) reduction, 15 shape features are calculated from the candidate regions. Then, they are given to the support vector machine; the candidate regions are classified either as true positive or FP. In the experiment, we compared the results of the proposed method and physician interpretation report using 200 images (63 architectural distortions) from a digital database of screening mammography. Experimental results indicate that our method may be effective to improve the performance of computer aided detection in mammography.

During mammographic acquisition, the breast is compressed between the breast support plate and the compression paddle to improve image quality and reduce dose, among other reasons. The applied force, which is measured by the imaging device, varies substantially, due to local guidelines, positioning, and breast size. Force measurements may not be very relevant though, because the amount of compression will be related to pressure rather than force. With modern image analysis techniques, the contact surface of the breast under compression can be determined and pressure can be computed retrospectively. In this study, we investigate if there is a relation between pressure applied to the breast during compression and screening performance.In a series of 113,464 screening exams from the Dutch breast cancer screening program we computed the compression pressure applied in the MLO projections of the right and left breasts. The exams were binned into five groups of increasing applied pressure, in such a way that each group contains 20 % of the exams. Thresholds were 7.68, 9.18, 10.71 and 12.81 kPa. Screening performance measures were determined for each group. Differences across the groups were investigated with a Pearson’s Chi Square test.It was found that PPV and the cancer detection rate vary significantly within the five groups (p = 0.001 and p = 0.011 respectively). The PPV was 25.4, 31.2, 32.7, 25.8 and 22.0 for the five groups with increasing pressure. The recall rate, false positive rate and specificity were not statistically significant from the expectation (p-values: 0.858, 0.088 and 0.094 respectively). Even though differences are not significant, there is a trend that the groups with a moderate pressure have a better performance compared to the first and last category.The results suggest that high pressure reduces detectability of breast cancer. The best screening results were found in the groups with a moderate pressure.

The mean glandular dose in mammography is evaluated via the normalized glandular dose coefficients (DgN), calculated via Monte Carlo simulations. The conversion from dose to the homogenous mixture to dose in the glandular tissue is made by considering an energy-dependent correction factor, G, which is the weighted mean of the energy absorption coefficients of adipose and glandular tissues. The authors implemented a GEANT4 code and evaluated, in the range 8−80 keV, the influence on the calculation of DgN values by (1) the method of G-weighting the dose, (2) the inclusion of bremsstrahlung radiation and (3) the energy threshold under which electrons are not tracked. The results for monochromatic DgN show that evaluating G retrospectively causes an underestimation up to 5 %, and that not considering bremsstrahlung or setting high electron energy cutoff may cause a bias up to 1 %, in the calculation of monochromatic DgN. These deviations may be negligible for polychromatic mammographic spectra.

In this work we evaluate an approach for breast density assessment of digital breast tomosynthesis (DBT) data using the central projection image. A total of 348 random cases (both FFDM CC and MLO views and DBT MLO views) were collected using a Siemens Mammomat Inspiration tomosynthesis unit at Unilabs, Malmö. The cases underwent both BI-RADS 5th Edition labeling by radiologists and automated volumetric breast density analysis (VBDA) by an algorithm. Preliminary results showed an observed agreement of 70 % (weighted Kappa, κ = 0.73) between radiologists and VBDA using FFDM images and 63 % (κ = 0.62) for radiologists and VBDA using DBT images. Comparison between densities for FFDM and DBT resulted in high correlation (r = 0.94) and an observed agreement of 72 % (κ = 0.76). The automated analysis is a promising approach using low dose central projection DBT images in order to get radiologist-like density ratings similar to results obtained from FFDM.

Scattering is one of the main issues left in planar mammography examinations, as it degrades the quality of the image and complicates the diagnostic process. Although widely used, anti-scatter grids have been found to be inefficient, increasing the dose delivered, the equipment price and not eliminating all the scattered radiation. Alternative scattering reduction methods, based on post-processing algorithms using Monte Carlo (MC) simulations, are being developed to substitute anti-scatter grids. Idealized detectors are commonly used in the simulations for the purpose of simplification. In this study, the scatter distribution of three detector geometries is analyzed and compared: Case 1 makes use of idealized detector geometry, Case 2 uses a scintillator plate and Case 3 uses a more realistic detector simulation, based on the structure of an indirect mammography X-ray detector. This paper demonstrates that common configuration simplifications may introduce up to 14 % of underestimation of the scatter in simulation results.

Our purpose was to investigate the influence of phantom and biological materials on a 3-component decomposition using dual-energy mammography protocol (3CB).Materials and Methods: A novel dual-energy 3CB mammography technique concludes in quantifying of the lipid, protein, and water thicknesses. The protocol was designed to be used on full-field digital mammography system by including an additional high-energy image with the clinical image. We study influence of calibration phantom and regression techniques on three component outputs. Two types of phantoms were used: solid water/wax/Delrin phantom and bovine phantom consisted of fat and lean muscle compartments. The linear and quadratic model equations were analyzed using linear and ridge regressions. The elaborated calibration protocol was applied to breast images with different compositions and sizes. In addition, the protocol was validated using cadaver breasts of known compositions.Results: We found that there were many negative values of protein components when we applied our solid water/wax/Delrin calibrations using 51 ROIs for clinical dual energy mammogram analysis. This behavior could be explained by potential over fitting and not exact correspondence of biological and phantom material. Creating a calibration related to bovine tissue provided higher accuracy and realizable thicknesses for clinical breast composition components, and achieved satisfactory results for cadaver breast compositions.Conclusion: Using a bovine calibration, the 3CB technique provides higher accuracy for lipid, water and protein compositional breast measurements than using plastic tissue equivalents alone.

High mammographic density is associated with reduced sensitivity of mammography. Recent changes in the BI-RADS density assessment address the potential for dense tissue to mask lesions, but the assessment remains qualitative and achieves only moderate agreement between radiologists. We have developed an automated, quantitative algorithm that generates a local detectability (d_L) map, which estimates the likelihood that a simulated lesion would be missed if present. The d_L map is computed by tessellating the mammogram into overlapping regions of interest, for which the detectability of a simulated lesion by a non-prewhitening model observer is calculated using local estimates of the noise power spectrum and volumetric breast density. The algorithm considers both the effects of loss of contrast due to density and the distracting appearance of density on lesion conspicuity.In previous work, it has been shown that the mean d_L from the maps are strongly correlated to detection performance by computerized and human readers in a controlled reader study. Here, we investigate how various statistical features of the d_L maps (gray-level histogram and co-occurrence features) are related to the diagnostic performance of mammography in a set of images comprised of 8 cancer cases that were mammographically occult and 40 cancer that were detected in screening mammography.

There are several methods to evaluate objectively the quality of a digital image. For digital mammography, objective quality assessment must be performed without references. In a previous study, the authors investigated the use of a normalized anisotropic quality index (NAQI) to assess mammography images blindly in terms of noise and spatial resolution. Since the NAQI is used as a quality metric, it must not be highly dependent on the breast anatomy. Thus, in this work, we analyze the NAQI behavior with different breast anatomies. A computerized system was used to synthesize 2,880 anthropomorphic breast phantom images with a realistic range of anatomical variations. The results show that NAQI is only marginally dependent on breast anatomy when images are acquired without degradation (<12 %). However, for realizations that simulate the acquisition process in digital mammography, the NAQI is more sensitive (33 %) to variations arising from quantum noise. Thus, NAQI can be used in clinical practice to assess mammographic image quality.

Purpose: To perform a virtual clinical trial study to assess the justification of the grid-less mammography acquisition mode with scatter correction software, as developed by Siemens Healthcare (PRIME mode). Materials and methods: The study was performed on a Siemens mammography unit using the conventional acquisition mode (system 1) and a second system used PRIME. Mean glandular doses (MGD) were compared from data of 5981 images. A paired t-test for all thickness groups (<29 mm, 30–49 mm, 50–70 mm, >69 mm) separately and combined had shown a significantly higher average MGD for system 1 (NON-PRIME) when compared to system 2 (PRIME), with an overall decrease of 11.7 %. The next phase in justification focused on detectability performance, in particular for screening applications. A dataset mimicking an enriched screened population was created by simulating previously developed anthropomorphic mass models and microcalcification clusters in 60 out of 100 normal mammograms of system 1 (NON-PRIME). The same physical lesions were then simulated into 60 out of 100 PRIME, normal mammograms. Care was taken to simulate each lesion model in matched mammograms PRIME-NON PRIME in terms of BI-RADS score, in a region with the same background glandularity (obtained after analysis with Volpara) and in a breast of the same thickness group. All images were visualized with ViewDEX software and four radiologists performed the free search detectability study. A JAFROC analysis was executed and detectability was quantified by means of the AUC. Results: Present approach allowed the realization of paired virtual clinical data sets starting from 200 normal mammograms. The results of all readers separately as well as combined showed approximately the same AUC for PRIME and NON-PRIME (0.57 vs 0.60), and the ANOVA analysis showed no statistical significant difference in detectability of the lesions between PRIME and NON-PRIME (p-value 0.36). The same result was found if the dataset was subdivided for both types of lesions: masses (p-value 0.88) and microcalcification clusters (p-value 0.33). Conclusion: Results state that the MGD is significantly lower in PRIME mode than with the conventional acquisition while lesion detectability remained constant for all four radiologists.

The aim of this work is to test whether a 3D structured phantom with simulated lesions can be used for performance evaluation of 2D digital mammography, as a step towards a multimodality phantom. A phantom, developed for breast tomosynthesis was therefore applied on 23 digital mammography systems. Ten images were acquired at the clinically used dose and for 11 systems also at half and double dose. The images were read in a four-alternative forced choice (4-AFC) paradigm by 5 readers. CDMAM phantom acquisitions were also performed. It was possible to calculate diameter thresholds of the simulated masses and microcalcifications that guarantee 62.5 % correct response. The results showed the expected sensitivity with mean glandular dose: detectability of microcalcifications improved with dose, whereas the detectability of masses was not affected. Systems of the same manufacturer and operated at similar doses had very similar detectability scores. Percentage correctly detected microcalcifications with average diameter 119 µm correlated with CDMAM based gold thickness thresholds. Present phantom, developed and tested for tomosynthesis, is also a good candidate for 2D mammography, suggesting its use for (future) benchmarking of at least two types of imaging systems.

Today, breast surgeons plan their procedures using pre-operatively placed metal clips or radioactive seeds and radiological images. These images show the breast in a positioning different from the one during surgery. We show a research prototype that eases the surgeon’s planning task by providing 3D visualizations based on the radiological images. With a FEM-based deformation simulation, we mimic the real surgical scenario. In particular, we have developed a ligament model that increases the robustness of a fully automatic prone-supine deformation simulation, and we have developed specific visualization methods to aid intra-operative breast lesion localization.

Measuring colocalization of multiple biomarkers may contribute to understanding tumor growth and progression. Traditionally, multiple biomarkers colocalization has been performed by processing multiple serial tissue sections. To provide true colocalization measures, we are investigating single section multiplexing techniques. Utilizing a fluorescent-based sequential stain and bleach system (SSB) we investigated multiplexing 8 markers in breast cancer tissue microarray sections. The experiments consisted of a 4 predictive biomarker panel (ER, PR, HER2, and Ki67) and markers to assist in the image processing. The goals included comparing the immunofluorescent signal with bright field single chromogen IHC scores, the measurements of tumor heterogeneity, and to discover technical challenges. Early results suggest that SSB signals correspond to traditional IHC staining. Additionally, our work has highlighted improvements to workflow and the staining process. We present a review of our progress and expectations for this technology.

Patient-specific surgical predictions of Breast Conserving Therapy, through mechano-biological simulations, could inform the shared decision making process between clinicians and patients by enabling the impact of different surgical options to be visualised. We present an overview of our processing workflow that integrates MR images and three dimensional optical surface scans into a personalised model. Utilising an interactively generated surgical plan, a multi-scale open source finite element solver is employed to simulate breast deformity based on interrelated physiological and biomechanical processes that occur post surgery. Our outcome predictions, based on the pre-surgical imaging, were validated by comparing the simulated outcome with follow-up surface scans of four patients acquired 6 to 12 months post-surgery. A mean absolute surface distance of 3.3 mm between the follow-up scan and the simulation was obtained.

Breast cancer is diagnosed by a combination of modalities. Measuring the elasto-mechanical properties of suspicious lesions, by e.g. ultrasound elastography, can help differentiate malignant from benign findings. Using a prototype Mechanical Imaging (MI) system as an adjunct to mammography, the aim of this study was to characterize tumors using MI and compare the readings to those from the contralateral breast. Thirteen bilateral MI sets from women with malignant breast lesions were included in this study, drawn from a larger set of 155 women recalled from screening. The results showed that mean lesion pressure was significantly greater than the mean pressure of the corresponding breast, 7.5 ± 7.0 kPa compared to 2.5 ± 1.6 kPa (P = 0.01). There was no evidence for a difference in mean pressure or standard deviation of the MI image between symptomatic and contralateral asymptomatic breasts (P = 0.24 and 0.68). The results support that it is possible to use MI to distinguish malignant cancers from normal breast tissue. Still, further investigations of the characteristics of benign lesions are necessary to ascertain the usefulness of the system.

This study compared mammographic density over time between women who developed breast cancer (cases) and women who did not (controls). Cases had an initial negative mammographic screen and another three years later when cancer was diagnosed. Cases were matched to three controls with two successive negative screens by age, year of mammogram, BMI, parity, menopausal status and HRT use. Mammographic density was measured by VolparaTM. There was a significant reduction in percentage density in the affected breast for cases (5.2 to 4.8 %, p < 0.001) and for the same matched breast in controls (4.9 to 4.5, p < 0.001). Similar results were found for the unaffected breast. After adjusting for density measures at the initial screen, case-control status was only significantly associated with fibroglandular volume in the unaffected breast (adjusted mean 45.8 cm3 in cases, 44.0 cm3 in controls, p = 0.008). The results suggest changes in mammographic density may be less important than initial mammographic density.

Breast cancer risk assessment is becoming increasingly important in clinical practice. It has been suggested that features that characterize mammographic texture are more predictive for breast cancer than breast density. Yet, strong correlation between both types of features is an issue in many studies. In this work we investigate a method to generate texture features and/or scores that are independent of breast density. The method is especially useful in settings where features are learned from the data itself. We evaluate our method on a case control set comprising 394 cancers, and 1182 healthy controls. We show that the learned density independent texture features are significantly associated with breast cancer risk. As such it may aid in exploring breast characteristics that are predictive of breast cancer irrespective of breast density. Furthermore it offers opportunities to enhance personalized breast cancer screening beyond breast density.

Aim: Understanding both normal mammographic appearance and how false positive (FP) errors occur is paramount to improving the efficiency and diagnostic accuracy of screening mammography services. While much of the focus of research is on increasing knowledge about the appearances and imaging of breast cancers, this study reports on findings where breast screen readers are asked to comment on past incorrect decisions by assigning a lexicon that best describes a known FP region. Method: Fifteen breast screen readers were given two tasks. The first was to assess nine normal screening cases which had attracted a high number of FP decisions in a test set of 60 cases in a previous study with 129 readers. In the second task, the 15 readers in this study, who were made aware that the nine cases were normal, were directed to view distinct regions of interest (ROI) that represented the FP markings from past readings in the blinded observer performance study. A list of descriptors derived from literature was used to assist readers to describe the mammographic appearance within those ROIs. Results: In the first task, readers identified breast density as the greatest difficulty in determining normality. In the second task, asymmetry of breast tissue and a suspicion of architectural distortion (AD) were the top two reasons our readers gave to explain the high number of past FP decisions. Additionally, our readers believed past FP decisions were less likely to reflect a suspicion of breast lesions or masses (second task). Conclusion: The classification of normal cases remains a challenging task, influenced by asymmetry and breast density. FP decisions may reflect a suspicion of AD and appear less related to suspicion of masses.

Two methods for estimation of location-dependent background tissue complexity (BTC) are proposed. The methods operate by calculating the lowest possible amplitude for which a small superimposed lesion remains visible at a given location in a mammogram: the higher BTC, the larger lesion insertion threshold amplitude. The visibility analysis is based on comparing a region of interest pre- and post-lesion using structural similarity metric (SSIM) in one method. The other proposed estimator is based on just noticeable difference (JND) notion Barten used in modeling contrast sensitivity function (we theorize that lesion detection is equivalent to detection of one cycle of a sinusoid). The proposed BTC estimators are evaluated by comparing them against the lesion insertion amplitude required for visibility set by a human observer. Our results indicate that both estimators correlate with each other (Spearman rank correlation coefficient r_s of 0.76) and outperform constant insertion amplitude in terms of correlation with perceived tissue complexity. The SSIM-based estimator has a higher correlation with the human observer over 24 locales that the estimators disagreed most or both predicted large BTC (r_s of 0.73, vs. 0.34 for JND-based estimator). The proposed estimators may be used to construct a BTC-aware model observer with applications such as optimization of contrast-enhanced medical imaging systems, and creation of an image dataset to match the characteristics of a given population.

A nation-wide survey of patient dose data was carried out, using data recorded in 2013–2015. Data from 32,000 women were collected. The average dose for oblique views, for DR systems, was 1.65 mGy for all women, and 1.35 mGy for 50–60 mm breasts. There was a wide range of doses for different systems, with the highest more than twice the dose of the lowest (2.03 mGy and 0.91 mGy respectively for the Hologic Dimensions and Philips MicroDose L30 systems, averaged over all breasts). Image quality, as indicated by the threshold gold thickness for 0.25 mm details, was better (0.21 µm) for the Hologic systems; for all the others it was practically the same (0.28 µm), although their doses to the average breast varied over a wide range.

This article presents the results of a pilot dose survey including fifty patients who underwent combined screening: full field digital mammography (FFDM) plus digital breast tomosynthesis (DBT). The study also aimed to demonstrate the different dosimetric outcome from using different glandularity assumptions and dosimetry methods. The mean glandular dose to each patient was computed using Dance’s method with UK glandularity assumption. The calculations were repeated using Wu/Boone’s method with the “50–50” breast assumption and the results compared to those using Dance’s method. For the typical breasts, the dose from combined examination was around 9.56 mGy: 4.26 mGy from two-view FFDM and 5.30 mGy from two-view DBT. Adopting UK glandularity assumption was believed to more realistically reflect the population dose. The comparison between Dance’s and Wu/Boone’s methods indicated that the latter tended to show lower dose values with mean differences of −3.6 % for FFDM and −5.5 % for DBT.

Clinical evaluation of dose reduction studies in x-ray breast imaging is problematic because it is difficult to justify imaging the same patient at a variety of radiation doses. One common alternative is to use simulation algorithms to manipulate a standard-dose exam to mimic reduced doses. Although there are several dose-reduction simulation methods for full-field digital mammography, the availability of similar methods for digital breast tomosynthesis (DBT) is limited. This work proposes a method for simulating dose reductions in DBT, based on the insertion of noise in a variance-stabilized domain. The proposed method has the advantage of performing signal-dependent noise injection without knowledge of the noiseless signal. We compared clinical low-dose DBT projections and reconstructed slices to simulated ones by means of power spectra, mean pixel values, and local standard deviations. The results of our simulations demonstrate low error (<5 %) between real and simulated images.

This paper investigates the use of mereotopological barcodes to help non-experts classify microcalcification clusters as either benign or malignant. When compared against classification using the microcalcification cluster segmentation maps, the use of barcodes is able to see a significant improvement in classification performance with the AUC significantly increasing ($$p < 0.01$$) from 0.62 for images to 0.82 for barcodes on the MIAS dataset. This shows that barcodes could prove useful to aid clinicians with interpreting and classifying mammographic microcalcifications.

Breast cancer is the most frequently diagnosed cancer in women. To date, the exact cause(s) of breast cancer still remains unknown. The most effective way to tackle the disease is early detection through breast screening programmes. Breast density is a well established image based risk factor. An accurate dense breast tissue segmentation can play a vital role in precise identification of women at risk, and determining appropriate measures for disease prevention. Fractal techniques have been used in many biomedical image processing applications with varying degrees of success. This paper describes a fractal inspired approach to mammographic tissue segmentation. A multiresolution stack representation and 3D histogram features (extended from 2D) are proposed. Quantitative and qualitative evaluation was performed including mammographic tissue segmentation and density classification. Results showed that the developed methodology was able to differentiate between breast tissue variations. The achieved density classification accuracy for 360 digital mammograms is 78 % based on the BI-RADS scheme. The developed fractal inspired approach in conjunction with the stack representation and 3D histogram features has demonstrated an ability to produce quality mammographic tissue segmentation. This in turn can be found useful in early breast cancer detection, risk-stratified screening, and aiding radiologists in the process of decision making prior to surgery and/or treatment.

Women that are diagnosed with breast cancer often undergo surgery to remove either the tumour and some of the surrounding tissue (lumpectomy) or the whole breast (mastectomy). After surgery, the excised tissue is sliced at the pathology department, where specimen radiographs of the slices are typically acquired. Representative parts of the tissue are then sampled for further processing, staining and examination under the microscope. The results of histopathological imaging are used for tumour characterisation. As the 3D structure of the specimen is inevitably lost during specimen handling, reconstructing a volume from individual specimen slices could facilitate the correlation of histology to radiological imaging. This work proposes a novel method for a whole specimen volume reconstruction and is validated on six mastectomy cases. We also demonstrate how these volumes can be used as a means to map multiple histology slides to a whole mastectomy image (MRI or CT).

Mammography, the most effective early breast cancer detection technique, is associated with the risk of missed lesions in dense breasts, and excessive X-ray exposure. Accurate estimations of glandularity and radiation dose are important during screening. We propose a novel, inexpensive method for accurate glandularity quantification using pixel values in clinical digital mammograms and X-ray exposure spectra. Glandularities were calculated for 314 mammograms in Japanese women, and the Dance formula c-factor was applied to estimate breast doses. To investigate the relationship between breast thickness and missed lesions, images were classified into four categories based on the rate of missed lesions, and correlated with breast thickness. Glandularity decreased with increasing compressed breast thickness, indicating that commonly used breast doses (assumed 50% glandularity) significantly overestimate thin breasts and underestimate thick breasts. The missed lesion rate was higher for thinner compressed breast thicknesses. Accurate glandularity estimation could thus promote individualized screening mammography.

Deep learning is a powerful tool in computer vision areas, but it is most effective when applied to large training sets. However, large dataset are not always available for medical images. In this study we proposed a new method to use deep neural network for near-term breast cancer risk analysis. In our data base, we have 420 cases with two sequential mammogram screenings, and half of the cases were diagnosed as positive in the second screening and the other half remained negative. Instead of using human designed features, we designed a deep neural network (DNN) with four pairs of convolution neural network and one fully connected layer. Every breast image were divided into 100 ROIs with 52 by 52 pixels, and each ROI were trained with the DNN individually, and the final predictions of each case were based on the overall risk scores of all the 100 ROIs. And the ROI based area under the curve (AUC) is 0.6982, and the case based AUC is 0.7173 using our proposed scheme. The results showed our proposed scheme is promising to apply deep learning algorithms in predicting near-term breast cancer risk with limited data size.

Computer aided diagnosis (CADx) schemes based on dual view mammograms are able to provide extra information compared to single view schemes. To explore an efficient and effective way for combining the information from different views, a new breast cancer risk analysis scheme was developed and tested in this study. 120 pairs of dual view mammograms from 120 women were used in this study. Three different groups of texture features and density features were extracted from both MLO view and CC view mammograms. The asymmetry score that measures the asymmetry levels of these two view mammograms was considered in our proposed scheme. 91 computational features on each view and 3 asymmetry measurements were computed and used for the proposed scheme. Three classifiers were used in our proposed scheme, one for each of the dual view mammograms, and the third one combined dual view scores with asymmetry measurements. The highest area under the curve (AUC) we obtained was 0.753 ± 0.039.

Computer aided detection (CADe) of breast cancer is mainly focused on monomodal applications. We propose an automated multimodal CADe approach, which uses patient-specific image registration of MRI and X-ray mammography to estimate the spatial correspondence of tissue structures. Then, based on the spatial correspondence, features are extracted from both MRI and X-ray mammography. As proof of principle, distinct regions of interest (ROI) were classified into normal and suspect tissue. We investigated the performance of different classifiers, compare our combined approach against a classification with MRI features only and evaluate the influence of the registration error. Using the multimodal information, the sensitivity for detecting suspect ROIs improved by 7 % compared to MRI-only detection. The registration error influences the results: using only datasets with a registration error below $$10\,mm$$, the sensitivity for the multimodal detection increases by 10 % to a maximum of 88 %, while the specificity remains constant. We conclude that automatically combining MRI and X-ray can enhance the result of a CADe system.

We retrospectively collected data on patient age, exposure factors, and compressed breast thickness (CBT) from 7,566 mammograms, which were obtained in the medio-lateral oblique projection over a 1-year period. The mean CBT was 31.7 mm, and was <30 mm in 44.8 % of cases. In 93.1 % of the mammograms with CBT 20-29 mm, tube voltage was 24 kV. In 196 mammograms exposed at 24 kV, the CBT was 29 mm and a maximum mAs value was over than 270 mAs. In order to evaluate the dose reduction, the tube loading was measured, using the semi-automatic AEC mode, for the mammography phantoms with a thickness of 10 to 30 mm while varying the tube voltage from 24 to 27 kV. The tube loading at 27 kV was approximately 50 % lower than that at 24 kV, the average glandular dose calculated could be reduced by about 20 %.

Recent advances in Computer-Aided Detection (CADe) for the automatic detection of clustered microcalcifications on mammograms show that cascade classifiers can compete with high-end commercial systems. In this paper, we introduce a deep cascade detector where the learning algorithm of each binary pixel classifier has been redesigned in the early stopping mechanism conventionally used to avoid overfitting to the training data. In this way, we strongly increase the number of features considered in each stage of the cascade (hence the term “deep”), yet we still benefit from the cascade framework by obtaining a very fast processing of mammograms (less than one second per image). We evaluated the proposed approach on a database of full-field digital mammograms; the experiments revealed a statistically significant improvement of deep cascade with respect to the traditional cascade framework. We also obtained statistically significantly higher performance than one of the most widespread commercial CADe systems, the Hologic R2CAD ImageChecker. Specifically, at the same number of false positives per image of R2CAD (0.21), the deep cascade detected 96 % of true lesions against the 90 % of R2CAD, whereas at the same lesion sensitivity of R2CAD (90 %), we obtained 0.05 false positives per image for the deep cascade against the 0.21 of R2CAD.

It has been shown that breast density and parenchymal patterns are important indicators in mammographic risk assessment. In addition, the accuracy of detecting abnormalities depends strongly on the structure and density of breast tissue. As such, mammographic parenchymal modelling and the related density estimation or classification are playing an important role in computer aided diagnosis. In this paper, we present a novel approach to the modelling of parenchymal tissue, which is directly linked to Tabar’s normal breast tissue representation and based on the multi-scale distribution of dark ellipses, and the complementary distribution of bright ellipses which represent dense tissue. Our initial evaluation is based on the full MIAS database. We provide analysis of the separation between the Birads density classes, which indicates significant differences and a way towards automatic Birads based density classification.

Slide quality is an important factor in pathology workflow and diagnosis. We examine the extent of quality variations in digitized hematoxylin-eosin (H&E) slides due to variations and errors in staining and/or scanning (e.g., out-of-focus blur & stitching). We propose two automatic quality estimators by adapting image quality assessment (IQA) methods that are originally developed for natural images. For the first estimator, we assume a gold-standard reference digital pathology slide is available. Quality of a given slide is estimated by comparing the slide to such a reference using a full-reference perceptual IQA method such as VIF (visual information fidelity) or SSIM (structural similarity metric). Our second estimator is based on IL-NIQE (integrated local natural image quality evaluator), a no-reference IQA, which we train using a set of artifact-free H&E high-power images (20× or 40×) from breast tissue. The first estimator (referenced) predicts marked quality reduction of images with simulated blurring as compared to the artifact-free originals used as references. The histograms of scores by the second estimator (no-reference) for images with artifact (blur, stitching, folded tissue, or air bubble artifacts) and for artifact-free images are highly separable. Moreover, the scores by the second estimator are correlated with the ratings given by a pathologist. We conclude that our approach is promising and further research is outlined for developing robust automatic quality estimators.

Analysis of invasive regions using breast magnetic resonance (MR) images plays an important role in diagnosis and decision-making regarding the treatment method. However, many images are obtained by MR imaging (MRI); development of an automated analysis method for breast tumours is desired. The main purpose of this study was to develop a novel method for automated analysis of the tumour region in breast MR images. First, early and late-subtraction images were obtained by subtracting early- and late-contrast-enhanced MR images, respectively, from the pre-contrast ones. Then, tumours in the images were enhanced based on the signal values of the normal mammary regions. Subsequently, using the level set method, a type of dynamic contour extraction, the outline of the tumour in the tumour-enhanced images was obtained. In order to evaluate the usefulness of the analysis method, we compared the tumour size listed in the interpretation report by a physician and analyzed the results obtained from the proposed method using clinical images from 10 cases. The mean absolute error of the size of tumours in all cases was less than 3.0 mm. These results indicate that the proposed method may be useful for the automated analysis of invasive breast tumours using breast MR images.

Virtual clinical trials (VCT) currently represent key tools for breast imaging optimisation, especially in two-dimensional planar mammography and digital breast tomosynthesis. Voxelised breast models are a crucial part of VCT as they allow the generation of synthetic image projections of breast tissue distribution. Therefore, realistic breast models containing an accurate representation of women breasts are needed. Current voxelised breast models show, in their compressed version, a very round contour which might not be representative of the entire population. This work pretends to develop an imaging framework, based on depth cameras, to investigate breast deformation during mammographic compression. Preliminary results show the feasibility of depth sensors for such task, however post-processing steps are needed to smooth the models. The proposed framework can be used in the future to produce more accurate compressed breast models, which will eventually generate more realistic images in VCT.

The aim of this study was to develop a tool to classify breast lesions using ultrasound elastography. Our dataset included a total of 78 patients enrolled for percutaneous biopsy of 85 breast lesions. These lesions were classified into three sonoelastographic scores, where scores of 1 and 2 were considered negative – soft and intermediate respectively; the score 3 was considered positive – hard. The visual classification of elastography performed by two radiologists was compared with our semi-automatic method. This classification aims to segment the red pixels found in the color elastography, quantify them and characterize the lesion by comparing the areas in red with the manually segmented lesion by the two radiologists. Our semi-automated technique had comparable performance to that of the two radiologists: sensitivity of 54.5 % and specificity of 90.5 %. The agreement kappa was greater than 0.8 for all observers. Thus, we concluded that the proposed method achieved a high rate of agreement between observers. In addition, the method presented high diagnostic specificity in classifying breast elastography images. By including more image features in the future, we expect our classifier can be use to standardize the classification of breast elastography.

In this paper, we propose a new method for automatic extraction of breast mass spicules in 2-D mammography. Spicules are abnormal curvilinear structures which characterize most of malignant breast masses. They are important features for discrimination between benign and malignant masses. In our method, the curvilinear structures are first approximated by line segments derived from localized Radon transforms; then, the Markov random field is used to take into account the local interactions via the contextual information between these segments. Finally, detection of the curvilinear structures that most likely correspond to spicules is performed using an a contrario framework. Validation of the approach was performed on a large dataset of spiculated masses which were selected from a public digital database; the results showed a high agreement with manually annotated mammograms.

Mammographic density is a strong risk factor for breast cancer. Volumetric breast density can be estimated from a digital mammogram by modelling the imaging process; this provides a more accurate assessment than subjective and 2D area-based methods. However, reliable density estimation in the uncompressed peripheral breast region and determination of compression paddle tilt are still open and challenging problems that affect the accuracy of measurement. Here we present a complete system that is able to perform thickness correction for both the compressed and uncompressed breast regions. The system was evaluated on a dataset of 208 mammograms, and compared with results from commercial software VolparaTM (version 1.5). The proposed method yielded Pearson correlation coefficients (PCC) of volumetric breast density (VBD) between left and right breasts of 0.88 (CC view) and 0.91 (MLO view). The PCC between VolparaTM VBD and our method is 0.93.

We have developed an efficient simulation of breast anatomy over a range of spatial scales, covering tissue details seen in both radiology and pathology images. The simulation is based on recursive partitioning using octrees, and is performed in two stages. First, the macro- and meso-scale anatomical features are simulated: breast outline, skin, and the matrix of tissue compartments and subcompartments, outlined by Cooper’s ligaments. These compartments are labeled as adipose or fibroglandular, according to the desired overall glandularity and the realistic distribution of dense tissue. Second, pathology images are generated to match selected region within the breast, by filling the region with simulated cells (adipocytes, ductal epithelium and myoepithelium, lymphocytes, and fibroblasts) and collagen fibers. Matched synthetic images can support discovery and virtual trials of image-based biomarkers for specific pathology findings. Our proof-of-concept is presented and further optimizations of the simulation discussed.

Automatic characterization of breast density can enable more personalized breast cancer screening work flows. In this work, we present a novel method to automatically characterize breast density in mammography images. Our method computes a volumetric density map and measures the relative volume of glandular tissue (VBD%). For critical cases when masking of small masses may be possible it additionally quantifies the masking effect of glandular tissue. VBD% and the masking risk combined provide a 4-point density score that correlates with the BI-RADS 5th edition guidelines. We evaluated our approach using a study with 32 radiologists and 2400 breast images (600 4-view FFDM exams). In a subset of 415 images identified as critical cases the accuracy to detect dense breasts (density categories c or d) increased as shown by the area under the curves (0.783 vs. 0.621). By taking masking risk into consideration our method provides a more comprehensive assessment of breast density.

Breast MRI to X-ray mammography registration usingpatient-specific biomechanical models is one challenging task in medical imaging. To solve this problem, the accurate knowledge about internal and external factors of the breast, such as internal tissues distribution, is needed for modelling a suitable physical behavior. In this work, we compare four different tissue segmentation algorithms, two intensity-based segmentation algorithms (Fuzzy C-means and Gaussian mixture model) and two improvements that incorporate spatial information (Kernelized Fuzzy C-means and Markov Random Fields, respectively), and analyze their effect to the multi-modal registration. The overall framework consists on using a density estimation software (Volpara$$^{TM}$$) to extract the glandular tissue from full-field digital mammograms, meanwhile, a biomechanical model is used to mimic the mammographic acquisition from the MRI, computing the glandular tissue traversed by the X-ray beam. Results with 40 patients show a high agreement between the amount of glandular tissue computed for each method.

The purpose of this work was to implement and evaluate the performance of a 3D Total Variation (TV) minimization filter for Poisson noise and apply it to 3D digital breast tomosynthesis (DBT) data. The value of Lagrange multiplier (λ) to be used in filter equation has a direct relationship with the results obtained. Some preliminary studies about λ values were done. A Mammographic Accreditation Phantom Model 156 was acquired and its biggest tumor-like mass and cluster of microcalcifications were used for image quality assessment. The proposed methodology was also tested with one clinical DBT data set.For 3D filter performance analysis: a reduction of 41.08 % and 38.60 % in 3D TV was achieved when a constant or variable λ value is used over slices, respectively. Either for constant or variable λ, the artifact spread function was improved, when compared to the unfiltered data. For the in-plane analysis: when constant λ is used, a reduction of 37.02 % in TV, an increase of 47.72 % in contrast to noise ratio (CNR) and a deterioration of 0.15 % in spatial resolution were obtained. For a variable λ, a reduction of 37.12 % in TV, an increase of 42.66 % in CNR and an improvement of 18.85 % in spatial resolution were achieved. A visual inspection of unfiltered and filtered clinical images demonstrates the quantitative values achieved with the phantom, where areas with higher noise level become smoother while preserving edges and details of the structures (about 43 % of TV reduction).Both quantitative and qualitative analysis performed in this study confirmed the relevance of this approach in improving image quality in DBT imaging.

This study investigates variations in mammographic density by ethnic group in women attending the NHS breast screening programme in Greater Manchester. Density was estimated using VolparaTM and QuantraTM. Data was analysed for 651 Asian/Asian British, 416 Black/Black British, 394 Jewish origin, 181 ‘Mixed’, 700 ‘Other’ and a random sample of 10,000 women who declared their ethnic origin as White (British or Irish). Age ranged from 46–84 years and mean BMI was 27.4 kg/m2. Fibroglandular volume (VolparaTM) was highest in women of Black/Black British origin (59.4 cm3) and lowest in Asian/Asian British women (47.9 cm3). After adjusting for a number of hormonal and other factors the magnitude of the difference between groups decreased, however, there were still a number of statistical differences between groups. Ethnic differences in mammographic density and personal factors may subsequently contribute to differences in breast cancer incidence.

The beginning of this century saw the development of simulation methods for the evaluation of breast imaging, motivated by the limitations of conventional clinical trials. This has led to the formation of AAPM Task Group on Virtual Tools for the Validation of 3D/4D X-ray Breast Imaging Systems (TG234), gathering researchers from academia, industry, and government, interested in the development, testing, and adoption of these tools. TG234 is currently finalizing its report. The report has been designed as an experiential guide through the steps of simulating breast anatomy, image acquisition, image interpretation, and analysis. TG234 activities include disseminating the idea of virtual clinical trials through numerous focused conference sessions and AAPM annual meeting symposia. This paper reflects our desire to initiate wider discussion about the future directions in the development of virtual tools for the design and evaluation of novel breast imaging systems.

Detectability of invasive cancerous lesions in mammography is diminished in breasts with dense and complex fibroglandular tissue. If masking were locally quantified in mammograms, radiologists could potentially use this information to clear the region using targeted adjuvant screening techniques. We present a method to quantify localized masking using a model observer to detect virtual objects of varying thickness and size convoluted into the clinical mammogram. Contrast detail curves are used to create an Image Quality Factor (IQF) at high resolution throughout the breast. We report on preliminary findings of how IQF is related to measures of breast density and textural complexity in a cohort of women who experienced screening detected and interval (masked) cancers. This measure of masking using a localized contrast detail curve approach should provide a means to target adjuvant screening resources for faster and more effective determination of cancer status for women with dense breast.

This study proposes a method to estimate low contrast detectability (LCD) applying a statistical method, based on the analysis of a uniform region. A dedicated test object was designed, made up of an acetate sheet equipped with a central uniform insert and an aluminium step wedge, allowing linear conversion from pixel values to millimeters of aluminium. A Matlab program for automated image analysis was developed. Phantom images were acquired on two different digital mammography systems. Reproducibility and sensitivity to exposure variations of the proposed method were investigated for different dose levels. Further the impact of scattering and attenuation on LCD was studied adding PMMA layers of variable thickness (2 to 7 cm) upon the acetate sheet during exposure in automatic exposure control modality. The statistical method turned out to be a reliable and rapid method for LCD evaluation. Applications include routine assessment of equipment performance for digital mammography systems.

This study aimed to determine whether correcting for observer variability alters estimations of breast cancer risk associated with mammographic density. A case control design examined the relationship between mammographic density, measured by visual analogue scales (VAS), and the risk of breast cancer after correcting for observer variability. Mammographic density was assessed by two observers and average scores (V2) were adjusted to correct for observer variability (V2_ad). Two case-control sets were identified: (i) breast cancer detected during screening at entry and (ii) breast cancer detected subsequently. Cases were matched to three controls. In the first case-control set the odds ratio for breast cancer was 4.6 (95 %CI 2.8–7.5) for the highest compared to the lowest quintile of V2, and was attenuated for V2_ad (OR 3.1, 95 %CI 1.9–4.8). Similar findings were observed for the second case-control set. Not adjusting for observer variability may lead to an overestimate of the risk of breast cancer.

Average volumetric breast density has been found to be associated with interval cancers. The association is believed to be partly due to mammographic masking. We asked if regional density may be a more sensitive descriptor of masking than average density. In this work, we propose a new method to identify high density regions based on calibrating pixel-level volumetric breast density to Breast Imaging Reporting and Data System (BI-RADS) Version 4 categories. Local breast density was measured using the single-energy X-ray absorptiometry (SXA) technique. In 583 women undergoing screening mammography, we found percent fibroglandular volume ranges that corresponded to each BI-RADS category: 0–4.9 % (BI-RADS 1), 5.0–18.2 % (BI-RADS 2), 18.3–48.9 % (BI-RADS 3) and 49.0–100 % (BI-RADS 4). Women with an average BI-RADS 1 category had 21840 pixels (0.014 mm) breast area considered high density (category 4) compared to 186469 pixels (0.014 mm) in women with average BI-RADS 4 category. We conclude that some women with low breast density still have regions of high density that may mask breast cancers. These scores and localized density colorized maps may better help radiologists in the decision to utilize secondary adjuvant screening than whole breast BI-RADs scores.

Approximately 75 % of biopsies performed in women with suspicious lesions are not found to have cancer. To reduce the number of false positive mammographic findings and the resulting biopsies, we have developed a combination of 3-compartment (lipid, water, protein) mammogram analysis and quantitative descriptors of lesions morphometry. The solution for lesion composition requires the calibration of dual energy mammogram attenuations to tissue equivalent standards. However, these phantom substitutes for actual breast tissue have inherent differences in energy-dependent X-ray characteristics that may lead to systematic errors in estimating composition. Here, we investigate the energy dependence of two biological materials (oil and water) to their phantom equivalents at different X-ray energies through both theoretical and empirical considerations. We first derived the relationship of the oil and water to the phantom materials at each energy and then compared to experimental measures. We found that the errors as large as 20 % in actual oil/water fraction when compared to the phantom materials at different energies. We conclude that calibrating breast composition to phantom materials at each dual-energy acquisition is not sufficient to ensure accuracy. A further basis transformation derived from bovine tissues greatly reduced these errors.

This work describes initial measurements with the CT subsystem of the assembled, fully-3D, hybrid SPECT-CT system for dedicated breast imaging. The hybrid system, designed for clinical breast imaging, consists of fully-flexible SPECT and CT subsystems, with each capable of 3D acquisition motions. The SPECT subsystem employs a 16 × 20 cm2 CZT detector with 2.5 mm pixellation, is capable of viewing into the chest wall in addition to imaging the complete breast volume, and has been extensively reported elsewhere. The polar tilting capability of the CT subsystem has marked improvement in volumetric sampling while eliminating cone beam artifacts due to the fully-3D acquisitions. The CT subsystem can also view into the chest wall, while delivering <5 mGy total dose, compared with a simple circular orbit breast CT. The CT subsystem consists of a 0.4 mm focal spot x-ray tube with a rotating 14° W-anode angle, and a 40 × 30 cm2 CsI(Tl) flat panel imager having 127 micron pixellation and 8.0 mm bezel edge, placed on opposing ends of the completely suspended gantry. A linear stage mechanism is used to tilt the suspended CT gantry up to ±15° in the polar directions about the 3D center of rotation; the SPECT system is nestled inside the suspended CT gantry, oriented perpendicular to the CT source-detector pair. Both subsystems rest on an azimuthal rotation stage enabling truncated spherical trajectories independently for each. Several simple and more complex 3D trajectories were implemented and characterized for the CT subsystem. Imaging results demonstrate that additional off-axis projection views of various geometric phantoms and intact cadaveric breast, facilitated by the polar tilting yield more complete breast-volume sampling and markedly improved iteratively reconstructed images, especially compared to simple circular orbit data. This is the first implementation of a hybrid SPECT-CT system with fully-3D positioning for the two subsystems, and could have various applications in diagnostic breast imaging.

Measurements of breast density have the potential to improve the efficiency and reduce the cost of screening mammography through personalized screening. Breast density has traditionally been evaluated from the dense area in a mammogram, but volumetric assessment methods, which measure the volumetric fraction of fibro-glandular tissue in the breast, are potentially more consistent and physically sound. The purpose of the present study is to evaluate a method for measuring the volumetric breast density using photon-counting spectral tomosynthesis. The performance of the method was evaluated using phantom measurements and clinical data from a small population ($$ n = 18 $$). The precision was determined to be 2.4 percentage points (pp) of volumetric breast density. Strong correlations were observed between contralateral ($$ R^{2} = 0.95 $$) and ipsilateral ($$ R^{2} = 0.96 $$) breast-density measurements. The measured breast density was anti-correlated to breast thickness, as expected, and exhibited a skewed distribution in the range [3.7 %, 55 %] and with a median of 18 %. We conclude that the method yields promising results that are consistent with expectations. The relatively high precision of the method may enable novel applications such as treatment monitoring.

A texture analysis aimed at finding correlations between textural descriptors and lesion diagnosis was applied to Contrast-Enhanced Digital Mammography (CEDM) subtracted images acquired under single-energy temporal subtraction modality using iodine-based contrast medium. The study, based on textural descriptors from Gray Level Co-occurrence Matrix (GLCM), included 68 CEDM images of 17 patients, 10 cancer and 7 benign, acquired 1 to 5 min after iodine injection. Seventeen GLCM descriptors were analyzed. Image processing consisted of geometric registration, logarithmic subtraction, and selection of regions-of-interest (adipose, glandular and lesion ROIs) by the radiologist. Results for lesion ROIs showed that homogeneity, normalized homogeneity, second-order inverse moment, energy and inverse variance were insensitive to the presence of iodine; a linear correlation existed between the sum mean and mean pixel value. Logistic regression showed that a linear combination of entropy and diagonal momentum discriminated between malignant and benign lesions with 79 % specificity, 93 % sensitivity and 87 % accuracy.

Dual-energy contrast-enhanced digital mammography (DE CE-DM) images the perfusion and vasculature of the breast using an iodinated contrast agent. High-energy (HE) and low-energy (LE) images of the breast are acquired; the DE image is obtained by a weighted logarithmic subtraction of the image pair. We hypothesized that the optimal DE subtraction weighting factor, w, is dependent on three parameters: breast thickness, kV, and filter material. We simulated the attenuation of x-rays through breasts of thicknesses ranging from 0.5 to 10 cm using different filter and kV combinations. The glandularity of the phantom for a given thickness was varied using different combinations of adipose and glandular tissues. We calculated the logarithm of the LE and HE signal intensities. For a given kV-filter pair, the signals decrease with increasing tissue thickness and glandularity. The DE weighting factor is thickness-dependent, and it decreases with increasing energy difference between the LE-HE kV pairs. These results facilitate the subtraction of tissue in the periphery of the breast, and aid in discriminating between contrast agent uptake in glandular tissue and subtraction artefacts.

Solitary, well-defined lesions are a common mammographic finding contributing more than 20 % of overall screening recalls. Discrimination of cystic from solid breast lesions therefore has the potential to reduce unnecessary recalls in mammography screening. A pre-clinical study, measuring the energy-dependent X-ray attenuation of tissue specimen and cystic fluid, revealed a measurable difference of the photon detection rate in the two energy bins of an energy-resolving photon-counting mammography system for these two tissue types. Based on these differences, a spectral lesion characterization algorithm has been developed, which estimates the lesion composition from spectral measurements in a lesion and a lesion-free reference region.In this work, we present a simulation study to estimate the dependence of this lesion characterization algorithm on various types of uncertainties including the biological variation of cyst fluid and tumor tissue, variations in the mammographic background texture, and errors in the spectral measurements. The simulation study uses the receiver operating characteristics (ROC) for the task of identifying solid lesions (‘positive result’) to predict an expected area under the curve (AUC) and the specificity at the 99 % sensitivity level for a simulated screening population. The results of this simulation study are compared to those of a recently published pilot study.

Primary goal of this study is to investigate the visibility of low-contrast details of different size on images obtained at conventional mammography unit, and at a monochromatic synchrotron radiation source, in absorption based and phase contrast imaging setups. For this purpose, three physical phantoms made of paraffin as a bulk material were used. They embedded various low contrast features. Single projection images were acquired with the GE Senographe mammography unit and at the beamline ID17, ESRF, Grenoble. Comparison of images showed that images obtained in a phase contrast mode have more visible details than the images acquired either in absorption mode at the synchrotron or at the conventional x-ray mammography unit. Analysis for δ and μ suggests that paraffin may be a suitable material for the manufacturing of tissue-mimicking phantoms dedicated to phase contrast applications. Results will be exploited in the development of a dedicated phantom for phase contrast imaging.

Digital mammography has limitations in sensitivity, in particular for patients with a dense breast. Phase contrast techniques (phase contrast mammography, PCM) might increase the tissue contrast for breast imaging. Propagation based PCM with a dedicated 0.1-mm-focal spot size mammography unit was investigated in past years, showing higher image quality in magnification PCM than in absorption based DM. In this work the authors investigated, using breast phantoms, the dependence of image quality on increasing mean glandular dose with a 0.007-mm-focal spot size W-anode microfocus X-ray tube. They compared PCM imaging (magnification M ≅ 2) to absorption based contact imaging (M ≅ 1) and then to phase retrieval for phase imaging, at low (40 kV) as well as high (80 kV) beam energy. Phase imaging shows higher image contrast for glandular masses and microcalcifications with MGD similar to one-view mammography. The phase contrast power spectrum assumes higher values than for absorption imaging. Possibility of dose reduction was suggested by the adoption of phase retrieval PCM.

We present the outcomes of combined feasibility studies carried out at Elettra and Australian Synchrotron to evaluate novel protocols for three-dimensional (3D) mammographic phase contrast imaging. A custom designed plastic phantom and some tissue samples have been studied at diverse resolution scales and experimental conditions. Several computed tomography (CT) reconstruction algorithms with different pre-processing and post-processing steps have been considered. Special attention was paid to the effect of phase retrieval on the diagnostic value of the reconstructed images. The images were quantitatively evaluated using objective quality indices in comparison with subjective assessments performed by three experienced radiologists and one pathologist. We show that the propagation-based phase-contrast imaging (PBI) leads to substantial improvement to the contrast-to-noise and to the intrinsic quality of the reconstructed CT images compared with conventional techniques as well as to an important reduction of the delivered doses, thus opening the way to clinical implementations.

In this paper, an optimization algorithm, utilizing a component measure of entropy, is developed for automatically tuning segmentation of mammograms by the Statistical Region Merging technique. The aim of this paper is to improve the mass segmentation in dense backgrounds. The proposed algorithm is tested on a database of 89 mammograms of which 41 have masses localized in dense background and 48 have masses in non-dense background. The algorithm performance is evaluated in conjunction with six standard enhancement techniques: Adjustable Histogram Equalization, Unsharp Masking, Neutrosophy based enhancement, standard CLAHE, Adaptive Clip Limit CLAHE based on standard deviation and Adaptive Clip Limit CLAHE based on standard entropy measure. For a comparison study, same experiments are performed using Fuzzy C-means Clustering technique. The experimental results show that the automatic tuning of SRM segmentation has the potential to produce an accurate segmentation of masses located in dense background while not compromising the performance on masses located in non-dense background.

In contrast-enhanced (CE) breast imaging, lesion contrast agent content may be diagnostic, relating to vascularization and tumour angiogenesis. However, factors in CE digital breast tomosynthesis (DBT) such as incomplete angular sampling, beam hardening and scatter can confound quantitative measurement. We propose system calibration to improve CEDBT quantitative potential. Fifteen projection images of an iodine calibration phantom, with various breast-equivalent compositions, thicknesses and lesion locations, were acquired over 40°, and reconstructed using SART. A separate breast-equivalent phantom with iodinated spherical lesions was imaged to test quantification accuracy. Excellent linearity between voxel signal and iodine concentration was achieved in the calibration data. In test phantom cases, lesion signal faithfully represented the true iodine concentration, but with wide confidence intervals for small lesions. While promising, it remains to be determined whether the iodine quantification accuracy is sufficient for lesion differentiation, and whether lesion shape, position, and non-uniform breast tissue would impair this estimation.

Nonuniform phantoms are needed in order to fully characterize the impact of anatomical structures on system performance in mammography and digital breast tomosynthesis (DBT). In this work, a new type of textured physical phantom is presented, compatible for use in both 2D and 3D applications. The breast phantom was first modeled analytically, and then fabricated using inkjet printing onto parchment paper and slide transparencies. A radiographic ink solution was synthesized with 350 mg/mL iohexol and pigmented ink. The effective linear attenuation coefficient (µ_eff) of the parchment paper alone (0.078 ± 0.003 mm−1) was found to be very close to that of a 70 % adipose, 30 % fibroglandular tissue mixture (0.078 ± 0.004 mm−1). The µeff of the parchment paper with iodine (0.010 ± 0.005 mm−1) was close to that of 100 % fibroglandular tissue (0.11 ± 0.004 mm−1). This new parchment and iodine phantom has strong potential for use in imaging studies.

Performance assessment of breast x-ray imaging systems through clinical imaging studies is expensive and may result in unreasonable high radiation doses to the patient. As an alternative, several research groups are investigating the potential of virtual clinical trials using realistic 3D breast texture models and simulated images from those models. This paper describes a mathematically defined solid 3D breast texture model based on the analysis of segmented clinical breast computed tomography images. The model employs stochastic geometry to mimic small and medium scale fibro-glandular and adipose tissue morphologies. Medium-scale morphology of each adipose compartment is simulated by a union of overlapping ellipsoids. The boundary of each ellipsoid consists of small Voronoi cells with average volume of 0.5 mm3, introducing a small-scale texture aspect. Model parameters were first empirically determined for almost entirely adipose breasts, scattered fibro-glandular dense breasts and heterogeneously dense breasts. Preliminary evaluation has shown that simulated mammograms and digital breast tomosynthesis images have a reasonable realistic visual appearance, depending though on simulated breast density. Statistical inference of model parameters from clinical breast computed tomography images for the variety of fibro-glandular and adipose tissue distributions observed in clinical images is ongoing.

Virtual clinical trials (VCTs) are increasingly being seen as a viable pre-clinical method for evaluation of imaging systems in breast cancer screening. The CR-UK funded OPTIMAM project is aimed at producing modelling tools for use in such VCTs. In the initial phase of the project, modelling tools were produced to simulate 2D-mammography and digital breast tomosynthesis (DBT) imaging systems. This paper elaborates on the new tools that have recently been developed for the current phase of the OPTIMAM project. These new additions to the framework include tools for simulating synthetic breast tissue, spiculated masses and variable-angle DBT systems. These tools are described in the paper along with the preliminary validation results. Four-alternative forced choice (4-AFC) type studies deploying these new tools are underway. The results of the ongoing 4AFC studies investigating minimum detectable contrast/size of masses/microcalcifications for different modalities and system designs are presented.

This work investigates the impact of advanced clinical displays on cancer detection in 2D digital mammograms using four-alternative-forced-choice (4AFC) and a dataset of images with inserted simulated lesions. Images were displayed on a standard monitor (Barco Coronis 5MP mammo) and an advanced monitor (Barco Coronis Uniti 12MP MDMC-12132). Ill-defined margin and spiculated mass models were inserted into mammographic regions of interest using a validated physics-based insertion framework. Experiments were conducted for mass size of 8–11 mm to 2–3 mm and density of 100 % to 70 % of glandular tissue with 142 trials per condition. Five medical physicists read the dataset on both monitors. Percentage correct (PC) of detected masses for average observer and 95 % confidence intervals were determined. Paired t-test and ANOVA analysis were performed. The observers had significantly better detection rates when the dataset was read on the advanced monitor compared to the standard monitor (3 % increase in overall PC, paired p-value $$=$$ 0.0076).