Skip to main content

Über dieses Buch

This book constitutes the refereed proceedings of the 6th International Workshop on Computational Methods and Clinical Applications for Musculoskeletal Imaging, MSKI 2018, held in conjunction with MICCAI 2018, in Granada, Spain, in September 2018.

The 13 workshop papers were carefully reviewed and selected for inclusion in this volume. Topics of interest include all major aspects of musculoskeletal imaging, for example: clinical applications of musculoskeletal computational imaging; computer-aided detection and diagnosis of conditions of the bones, muscles and joints; image-guided musculoskeletal surgery and interventions; image-based assessment and monitoring of surgical and pharmacological treatment; segmentation, registration, detection, localization and visualization of the musculoskeletal anatomy; statistical and geometrical modeling of the musculoskeletal shape and appearance; image-based microstructural characterization of musculoskeletal tissue; novel techniques for musculoskeletal imaging.



Automated Recognition of Erector Spinae Muscles and Their Skeletal Attachment Region via Deep Learning in Torso CT Images

Erector spinae muscle (ESM) is an important muscle in the torso region. Changes of sizes, shapes and densities in the cross section of the spinal column muscles have been found in chronic low back pain, degenerative lumbar sclerosis and chronic obstructive pulmonary disease. However, the image features of the ESM are measured manually by the physician. Therefore, automatic recognition in three dimensions (3D) not only for the limited two-dimensional (2D) section but also for the whole ESM is required. In this study, we realize automatic recognition of the ESMs and its attachment region on the skeleton using a 2D deep convolutional neural network. Each cross section of the 3D computed tomography (CT) image is input as a 2D image to the fully convolutional network. Then, the obtained result is reconstructed into a 3D image to obtain the recognition result of the ESM and its attachment region on the skeleton. ESM and attached area are extracted manually from the CT images of 11 cases and used for evaluation. In the experiments, automatic recognition was performed for each case using the leave-one-out method. The mean recognition accuracy of ESM and attached area was \(89.9\%\) and \(65.5\%\), respectively for the Dice coefficient. In this study, although there is over-extraction in the recognition of the attachment region, the initial region has been acquired successfully and it is the first study to simultaneously recognize the ESMs and its attachment region on the skeleton.
Naoki Kamiya, Masanori Kume, Guoyan Zheng, Xiangrong Zhou, Hiroki Kato, Huayue Chen, Chisako Muramatsu, Takeshi Hara, Toshiharu Miyoshi, Masayuki Matsuo, Hiroshi Fujita

Fully Automatic Teeth Segmentation in Adult OPG Images

This work addresses the problem of segmenting teeth in panoramic dental images. Random forest regression voting constrained local models were applied firstly to locate the mandible and the approximate pose of each tooth, and secondly to locate the full outline of each individual tooth. An automatically computed quality-of-fit measure was proposed to identify missing teeth. The system was evaluated using 346 manually annotated images containing adult-stage mandibular teeth. Encouraging results were achieved for detecting missing teeth. The system achieved state-of-the-art performance in locating the outline of present teeth with a median point-to-curve error of 0.2 mm for each of the teeth.
Nicolás Vila Blanco, Timothy F. Cootes, Claudia Lindner, Inmaculada Tomás Carmona, Maria J. Carreira

Fully Automatic Planning of Total Shoulder Arthroplasty Without Segmentation: A Deep Learning Based Approach

We present a method for automatically determining the position and orientation of the articular marginal plane (AMP) of the proximal humerus in computed tomography (CT) images without segmentation or hand-crafted features. The process is broken down into 3 stages. Stage 1 determines a coarse estimation of the AMP center by sampling patches over the entire image and combining predictions with a novel kernel density estimation method. Stage 2 utilizes the estimate from stage 1 to focus on a smaller sampling region and operates at a higher images resolution to obtain a refined prediction of the AMP center. Stage 3 focuses patch sampling on the region around the center obtained at stage 2 and regresses the tip of a vector normal to the AMP which yields the orientation of the plane. The system was trained and evaluated on 27 upper arm CTs. In a 4-fold cross-validation the mean error in estimating the AMP center was \(1.30\,{\pm }\,0.65\) mm and the angular error for estimating the normal vector was \(4.68\,{\pm }\,2.84^\circ \).
Paul Kulyk, Lazaros Vlachopoulos, Philipp Fürnstahl, Guoyan Zheng

Deep Volumetric Shape Learning for Semantic Segmentation of the Hip Joint from 3D MR Images

This paper addresses the problem of segmentation of the hip joint including both the acetabulum and the proximal femur in three-dimensional magnetic resonance images. We propose a fully convolutional volumetric auto encoder that learns a volumetric representation from manual segmentation in order to regularize the segmentation results obtained from a fully convolutional network. We further introduce a super resolution network to improve the segmentation accuracy. Comprehensive results obtained from 24 patient data demonstrated the effectiveness of the proposed framework.
Guodong Zeng, Guoyan Zheng

Pelvis Segmentation Using Multi-pass U-Net and Iterative Shape Estimation

In this report, an automatic method for segmentation of the pelvis in three-dimensional (3D) computed tomography (CT) images is proposed. The method is based on a 3D U-net which has as input the 3D CT image and estimated volumetric shape models of the targeted structures and which returns the probability maps of each structure. During training, the 3D U-net is initially trained using blank shape context inputs to generate the segmentation masks, i.e. relying only on the image channel of the input. The preliminary segmentation results are used to estimate a new shape model, which is then fed to the same network again, with the input images. With the additional shape context information, the U-net is trained again to generate better segmentation results. During the testing phase, the input image is fed through the same 3D U-net multiple times, first with blank shape context channels and then with iteratively re-estimated shape models. Preliminary results show that the proposed multi-pass U-net with iterative shape estimation outperforms both 2D and 3D conventional U-nets without the shape model.
Chunliang Wang, Bryan Connolly, Pedro Filipe de Oliveira Lopes, Alejandro F. Frangi, Örjan Smedby

Bone Adaptation as Level Set Motion

Bone microarchitecture is constantly adapting to environmental and mechanical factors. Changes in bone density and structure can lead to an increase in fracture risk. Computational modeling of bone adaptation may provide insight into mitigating aging and preventing disease. In this paper, the adaptation of bone is modeled as a curve evolution problem. Curves can be evolved according to the level set method. The level set method models basic bone physiology by adapting bone according to appositional growth following a trajectory in time with a natural definition of homeostasis. A novel curvature based bone adaptation algorithm is presented for modeling bone atrophy. The algorithm is shown to be weakly equivalent to simulated bone atrophy. These results generalize surface-driven and strain-driven models of bone adaptation using a surface remodeling force. Physiological signals (hormones, mechanical strain, etc.) can be directly integrated into this surface remodeling force. Remodeling can be naturally restricted around foreign bodies (such as modeling adaptation around a surgical screw). Future work aims to identify the surface remodeling force from longitudinal image data.
Bryce A. Besler, Leigh Gabel, Lauren A. Burt, Nils D. Forkert, Steven K. Boyd

Landmark Localisation in Radiographs Using Weighted Heatmap Displacement Voting

We propose a new method for fully automatic landmark localisation using Convolutional Neural Networks (CNNs). Training a CNN to estimate a Gaussian response (“heatmap”) around each target point is known to be effective for this task. We show that better results can be obtained by training a CNN to predict the offset to the target point at every location, then using these predictions to vote for the point position. We show the advantages of the approach, including those of using a novel loss function and weighting scheme. We evaluate on a dataset of radiographs of child hips, including both normal and severely diseased cases. We show the effect of varying the training set size. Our results show significant improvements in accuracy and robustness for the proposed method compared to a standard heatmap prediction approach and comparable results with a traditional Random Forest method.
Adrian K. Davison, Claudia Lindner, Daniel C. Perry, Weisang Luo, Timothy F. Cootes

Perthes Disease Classification Using Shape and Appearance Modelling

We propose to use statistical shape and appearance modelling to classify the proximal femur in hip radiographs of children into Legg-Calvé-Perthes disease and healthy. Legg-Calvé-Perthes disease affects the femoral head with avascular necrosis, which causes large shape deformities during the growth-stage of the child. Further, the dead or dying bone of the femoral head is prominent visually in radiographic images, leading to a distinction between healthy bone and bone where necrosis is present. Currently, there is little to no research into analysing the shape and appearance of hips affected by Perthes disease from radiographic images. Our research demonstrates how the radiographic shape, texture and overall appearance of a proximal femur affected by Perthes disease differs and how this can be used for identifying cases with the disease. Moreover, we present a radiograph-based fully automatic Perthes classification system that achieves state-of-the-art results with an area under the receiver operator characteristic (ROC) curve of 98%.
Adrian K. Davison, Timothy F. Cootes, Daniel C. Perry, Weisang Luo, Claudia Lindner

Deep Learning Based Rib Centerline Extraction and Labeling

Automated extraction and labeling of rib centerlines is a typically needed prerequisite for more advanced assisted reading tools that help the radiologist to efficiently inspect all 24 ribs in a computed tomography (CT) volume. In this paper, we combine a deep learning-based rib detection with a dedicated centerline extraction algorithm applied to the detection result for the purpose of fast, robust and accurate rib centerline extraction and labeling from CT volumes. More specifically, we first apply a fully convolutional neural network to generate a probability map for detecting the first rib pair, the twelfth rib pair, and the collection of all intermediate ribs. In a second stage, a newly designed centerline extraction algorithm is applied to this multi-label probability map. Finally, the distinct detection of first and twelfth rib separately, allows to derive individual rib labels by simple sorting and counting the detected centerlines. We applied our method to CT volumes with an isotropic voxel spacing of 1.5 mm from 113 patients which included a variety of different challenges and achieved a mean centerline accuracy of 0.723 mm with respect to manual centerline annotations. The presented approach can be applied to similar tracing problems, such as detecting the spinal column centerline.
Matthias Lenga, Tobias Klinder, Christian Bürger, Jens von Berg, Astrid Franz, Cristian Lorenz

Automatic Detection of Wrist Fractures From Posteroanterior and Lateral Radiographs: A Deep Learning-Based Approach

We present a system that uses convolutional neural networks (CNNs) to detect wrist fractures (distal radius fractures) in posterioanterior and lateral radiographs. The proposed system uses random forest regression voting constrained local model to automatically segment the radius. The resulting automatic annotation is used to register the object across the dataset and crop patches. A CNN is trained on the registered patches for each view separately. Our automatic system outperformed existing systems with a performance of 96% (area under receiver operating characteristic curve) on cross-validation experiments on a dataset of 1010 patients, half of them with fractures.
Raja Ebsim, Jawad Naqvi, Timothy F. Cootes

Bone Reconstruction and Depth Control During Laser Ablation

Cutting bones using laser light has been studied by several groups over the last decades. Yet, the risk of cutting nerves or soft tissues behind the bone is still an untackled problem. When performing tissue ablation such as bone, an acoustic signal is emitted. This paper presents a numerical framework that takes advantage of this acoustic signal to reconstruct not only the structure of the bone but also estimates the current cut position and depth. We employ an inverse problems approach to estimate the bone structure followed by an optimal control step to localize the position and depth of the signal source, i.e. the position of the cut. Besides the methodological description we also present numerical simulations in two dimensions with realistic mixed soft- and hard-tissue objects.
Uri Nahum, Azhar Zam, Philippe C. Cattin

Automated Dynamic 3D Ultrasound Assessment of Developmental Dysplasia of the Infant Hip

Dynamic two-dimensional sonography of the infant hip is a commonly used procedure for developmental dysplasia of the hip (DDH) screening by many clinicians. It however has been found to be unreliable with some studies reporting associated misdiagnosis rates of up to \(29\%\). Aiming to improve reliability of diagnosis and to help in standardizing diagnosis across different raters and health-centers, we present a preliminary automated method for assessing hip instability using three-dimensional (3D) dynamic ultrasound (US). To quantify hip assessment, we propose the use of femoral head coverage variability (\(\varDelta FHC_{3D}\)) within US volumes collected during a dynamic scan which uses phase symmetry features to approximate the vertical cortex of the ilium and a random forest classifier to identify the approximate location of the femoral head. We measure the change in \(FHC_{3D}\) across US scans of the hip acquired under posterior stress vs. rest as maneuvered during a 3D dynamic assessment. Our findings on 38 hips from 19 infants scanned by one orthopedic surgeon and two radiology technicians suggests the proposed \(\varDelta FHC_{3D}\) may provide a good degree of repeatability with an average test-retest intraclass correlation measure of 0.70 (\(95\%\) confidence interval: 0.35 to 0.87, \(F(21,21)\,{=}\,7.738\), \(p\,{<}\,0.001\)). This suggests that our 3D dynamic dysplasia metric may prove valuable in improving reliability in diagnosing hip laxity due to DDH, which may lead to a more standardized DDH assessment with better diagnostic accuracy. The long-term significance of this approach to evaluating dynamic assessments may lie in increasing early diagnostic sensitivity in order to prevent dysplasia remaining undetected prior to manifesting itself in early adulthood joint disease.
Olivia Paserin, Kishore Mulpuri, Anthony Cooper, Antony J. Hodgson, Rafeef Garbi

Automated Measurement of Pelvic Incidence from X-Ray Images

One of the most important parameters of sagittal pelvic alignment is the pelvic incidence (PI), which is commonly measured from sagittal X-ray images of the pelvis as the angle between the line connecting the midpoint of the femoral head centers with the center of the sacral endplate, and the line orthogonal to the sacral endplate. In this paper, we present the results of a fully automated measurement of PI from X-ray images that is based on the deep learning technologies. In each sagittal X-ray image of the pelvis, regions of interest (sacral endplate and both femoral heads) are first automatically defined, and then landmarks are detected within these regions, i.e. the anterior edge, the center and the posterior edge of the sacral endplate that define the line of the sacral endplate inclination, and the centers of both femoral heads with the corresponding midpoint representing the hip axis. From the hip axis, and the line along the sacral endplate and its center, PI is computed. Measurements were performed on X-ray pelvic images from 38 subjects (15 males/23 females; mean age 71.1 years), and statistical analysis of reference manual and fully automated measurements revealed a relatively good agreement, with the mean absolute difference ± standard deviation of \(5.1\,{\pm }\,4.4^\circ \) and Pearson correlation coefficient of \(R\,=\,0.82\) (p-value below \(10^{-6}\)), with the paired t-test revealing no statistically significant differences (p-value above 0.05). The differences between reference manual and fully automated measurements were within the repeatability and reliability of manual measurements, indicating that PI can be accurately determined by the proposed fully automated approach.
Robert Korez, Michael Putzier, Tomaž Vrtovec


Weitere Informationen

Premium Partner