Skip to main content

Über dieses Buch

The first International Workshop on Machine Learning in Medical Imaging, MLMI 2010, was held at the China National Convention Center, Beijing, China on Sept- ber 20, 2010 in conjunction with the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) 2010. Machine learning plays an essential role in the medical imaging field, including image segmentation, image registration, computer-aided diagnosis, image fusion, ima- guided therapy, image annotation, and image database retrieval. With advances in me- cal imaging, new imaging modalities, and methodologies such as cone-beam/multi-slice CT, 3D Ultrasound, tomosynthesis, diffusion-weighted MRI, electrical impedance to- graphy, and diffuse optical tomography, new machine-learning algorithms/applications are demanded in the medical imaging field. Single-sample evidence provided by the patient’s imaging data is often not sufficient to provide satisfactory performance; the- fore tasks in medical imaging require learning from examples to simulate a physician’s prior knowledge of the data. The MLMI 2010 is the first workshop on this topic. The workshop focuses on major trends and challenges in this area, and works to identify new techniques and their use in medical imaging. Our goal is to help advance the scientific research within the broad field of medical imaging and machine learning. The range and level of submission for this year's meeting was of very high quality. Authors were asked to submit full-length papers for review. A total of 38 papers were submitted to the workshop in response to the call for papers.



Fast Automatic Detection of Calcified Coronary Lesions in 3D Cardiac CT Images

Even with the recent advances in multidetector computed tomography (MDCT) imaging techniques, detection of calcified coronary lesions remains a highly tedious task. Noise, blooming and motion artifacts etc. add to its complication. We propose a novel learning-based, fully automatic algorithm for detection of calcified lesions in contrast-enhanced CT data. We compare and evaluate the performance of two supervised learning methods. Both these methods use rotation invariant features that are extracted along the centerline of the coronary. Our approach is quite robust to the estimates of the centerline and works well in practice. We are able to achieve average detection times of 0.67 and 0.82 seconds per volume using the two methods.
Sushil Mittal, Yefeng Zheng, Bogdan Georgescu, Fernando Vega-Higuera, Shaohua Kevin Zhou, Peter Meer, Dorin Comaniciu

Automated Intervertebral Disc Detection from Low Resolution, Sparse MRI Images for the Planning of Scan Geometries

Robust and accurate identification of intervertebral discs from low resolution, sparse MRI scans is essential for the automated scan planning of the MRI spine scan. This paper presents a graphical model based solution for the detection of both the positions and orientations of intervertebral discs from low resolution, sparse MRI scans. Compared with the existing graphical model based methods, the proposed method does not need a training process using training data and it also has the capability to automatically determine the number of vertebrae visible in the image. Experiments on 25 low resolution, sparse spine MRI data sets verified its performance.
Xiao Dong, Huanxiang Lu, Yasuo Sakurai, Hitoshi Yamagata, Guoyan Zheng, Mauricio Reyes

Content-Based Medical Image Retrieval with Metric Learning via Rank Correlation

A novel content-based medical image retrieval method with metric learning via rank correlation is proposed in this paper. A new rank correlation measure is proposed to learn a metric encoding the pairwise similarity between images via direct optimization. Our method has been evaluated with a large population-based dataset composed of 5000 slit-lamp images with different nuclear cataract severities. Experimental results and statistical analysis demonstrate the superiority of our method over several popular metric learning methods in content-based slit-lamp image retrieval.
Wei Huang, Kap Luk Chan, Huiqi Li, Joo Hwee Lim, Jiang Liu, Tien Yin Wong

A Hyper-parameter Inference for Radon Transformed Image Reconstruction Using Bayesian Inference

We propose an hyper-parameter inference method in the manner of Bayesian inference for image reconstruction from Radon transformed observation which often appears in the computed tomography. Hyper-parameters are often introduced in Bayesian inference to control the strength ratio between prior information and the fidelity to the observation. Since the quality of the reconstructed image is influenced by the estimation accuracy of these hyper-parameters, we apply Bayesian inference into the filtered back projection (FBP) reconstruction method with hyper-parameters inference, and demonstrate that estimated hyper-parameters can adapt to the noise level in the observation automatically.
Hayaru Shouno, Masato Okada

Patch-Based Generative Shape Model and MDL Model Selection for Statistical Analysis of Archipelagos

We propose a statistical generative shape model for archipelago-like structures. These kind of structures occur, for instance, in medical images, where our intention is to model the appearance and shapes of calcifications in x-ray radio graphs. The generative model is constructed by (1) learning a patch-based dictionary for possible shapes, (2) building up a time-homogeneous Markov model to model the neighbourhood correlations between the patches, and (3) automatic selection of the model complexity by the minimum description length principle. The generative shape model is proposed as a probability distribution of a binary image where the model is intended to facilitate sequential simulation. Our results show that a relatively simple model is able to generate structures visually similar to calcifications. Furthermore, we used the shape model as a shape prior in the statistical segmentation of calcifications, where the area overlap with the ground truth shapes improved significantly compared to the case where the prior was not used.
Melanie Ganz, Mads Nielsen, Sami Brandt

Prediction of Dementia by Hippocampal Shape Analysis

This work investigates the possibility of predicting future onset of dementia in subjects who are cognitively normal, using hippocampal shape and volume information extracted from MRI scans. A group of 47 subjects who were non-demented normal at the time of the MRI acquisition, but were diagnosed with dementia during a 9 year follow-up period, was selected from a large population based cohort study. 47 Age and gender matched subjects who stayed cognitively intact were selected from the same cohort study as a control group. The hippocampi were automatically segmented and all segmentations were inspected and, if necessary, manually corrected by a trained observer. From this data a statistical model of hippocampal shape was constructed, using an entropy-based particle system. This shape model provided the input for a Support Vector Machine classifier to predict dementia. Cross validation experiments showed that shape information can predict future onset of dementia in this dataset with an accuracy of 70%. By incorporating both shape and volume information into the classifier, the accuracy increased to 74%.
Hakim C. Achterberg, Fedde van der Lijn, Tom den Heijer, Aad van der Lugt, Monique M. B. Breteler, Wiro J. Niessen, Marleen de Bruijne

Multi-Class Sparse Bayesian Regression for Neuroimaging Data Analysis

The use of machine learning tools is gaining popularity in neuroimaging, as it provides a sensitive assessment of the information conveyed by brain images. In particular, finding regions of the brain whose functional signal reliably predicts some behavioral information makes it possible to better understand how this information is encoded or processed in the brain. However, such a prediction is performed through regression or classification algorithms that suffer from the curse of dimensionality, because a huge number of features (i.e. voxels) are available to fit some target, with very few samples (i.e. scans) to learn the informative regions. A commonly used solution is to regularize the weights of the parametric prediction function. However, model specification needs a careful design to balance adaptiveness and sparsity. In this paper, we introduce a novel method, Multi − Class Sparse Bayesian Regression(MCBR), that generalizes classical approaches such as Ridge regression and Automatic Relevance Determination. Our approach is based on a grouping of the features into several classes, where each class is regularized with specific parameters. We apply our algorithm to the prediction of a behavioral variable from brain activation images. The method presented here achieves similar prediction accuracies than reference methods, and yields more interpretable feature loadings.
Vincent Michel, Evelyn Eger, Christine Keribin, Bertrand Thirion

Appearance Normalization of Histology Slides

This paper presents a method for automatic color and intensity normalization of digitized histology slides stained with two different agents. In comparison to previous approaches, prior information on the stain vectors is used in the estimation process, resulting in improved stability of the estimates. Due to the prevalence of hematoxylin and eosin staining for histology slides, the proposed method has significant practical utility. In particular, it can be used as a first step to standardize appearances across slides, that is very effective at countering effects due to differing stain amounts and protocols, and to slide fading. The approach is validated using synthetic experiments and 13 real datasets.
Marc Niethammer, David Borland, J. S. Marron, John Woosley, Nancy E. Thomas

Parallel Mean Shift for Interactive Volume Segmentation

In this paper we present a parallel dynamic mean shift algorithm based on path transmission for medical volume data segmentation. The algorithm first translates the volume data into a joint position-color feature space subdivided uniformly by bandwidths, and then clusters points in feature space in parallel by iteratively finding its peak point. Over iterations it improves the convergent rate by dynamically updating data points via path transmission and reduces the amount of data points by collapsing overlapping points into one point. The GPU implementation of the algorithm can segment 256x256x256 volume in 6 seconds using an NVIDIA GeForce 8800 GTX card for interactive processing, which is hundreds times faster than its CPU implementation. We also introduce an interactive interface to segment volume data based on this GPU implementation. This interface not only provides the user with the capability to specify segmentation resolution, but also allows the user to operate on the segmented tissues and create desired visualization results.
Fangfang Zhou, Ying Zhao, Kwan-Liu Ma

Soft Tissue Discrimination Using Magnetic Resonance Elastography with a New Elastic Level Set Model

Magnetic resonance elastography (MRE) noninvasively images the propagation of mechanical waves within soft tissues. The elastic properties of soft tissues can then be quantified from MRE wave snapshots. Various algorithms have been proposed to obtain their inversion for soft tissue elasticity. Anomalies are assumed to be discernible in the elasticity map. We propose a new elastic level set model to directly detect and track abnormal soft tissues in MRE wave images. It is derived from the Mumford-Shah functional, and employs partial differential equations for function modeling and smoothing. This level set model can interpret MRE wave images without elasticity reconstruction. The experimental results on synthetic and real MRE wave images confirm its effectiveness for soft tissue discrimination.
Bing Nan Li, Chee Kong Chui, Sim Heng Ong, Toshikatsu Washio, Tomokazu Numano, Stephen Chang, Sudhakar Venkatesh, Etsuko Kobayashi

Fast and Automatic Heart Isolation in 3D CT Volumes: Optimal Shape Initialization

Heart isolation (separating the heart from the proximity tissues, e.g., lung, liver, and rib cage) is a prerequisite to clearly visualize the coronary arteries in 3D. Such a 3D visualization provides an intuitive view to physicians to diagnose suspicious coronary segments. Heart isolation is also necessary in radiotherapy planning to mask out the heart for the treatment of lung or liver tumors. In this paper, we propose an efficient and robust method for heart isolation in computed tomography (CT) volumes. Marginal space learning (MSL) is used to efficiently estimate the position, orientation, and scale of the heart. An optimal mean shape (which optimally represents the whole shape population) is then aligned with detected pose, followed by boundary refinement using a learning-based boundary detector. Post-processing is further exploited to exclude the rib cage from the heart mask. A large-scale experiment on 589 volumes (including both contrasted and non-contrasted scans) from 288 patients demonstrates the robustness of the approach. It achieves a mean point-to-mesh error of 1.91 mm. Running at a speed of 1.5 s/volume, it is at least 10 times faster than the previous methods.
Yefeng Zheng, Fernando Vega-Higuera, Shaohua Kevin Zhou, Dorin Comaniciu

Relation-Aware Spreadsheets for Multimodal Volume Segmentation and Visualization

Multimodal volume data commonly found in medical imaging applications present both opportunities and challenges to segmentation and visualization tasks. This paper presents a user directed volume segmentation system. Through a spreadsheets interface, the user can interactively examine and refine segmentation results obtained from automatic clustering. In addition, the user can isolate or highlight a feature of interest in a volume based on different modalities, and see the corresponding segmented results. Our system is easy to use since the preliminary segmentation results are organized and presented to the user in a relation-aware fashion based on the spatial relations between the segmented regions. We demonstrate this system using two multimodal datasets.
Lin Zheng, Yingcai Wu, Kwan-liu Ma

A Bayesian Learning Application to Automated Tumour Segmentation for Tissue Microarray Analysis

Tissue microarray (TMA) is a high throughput analysis tool to identify new diagnostic and prognostic markers in human cancers. However, standard automated method in tumour detection on routine histochemical images for TMA construction is under developed. This paper presents a MRF based Bayesian learning system for automated tumour cell detection in routine histochemical virtual slides to assist TMA construction. The experimental results show that the proposed method is able to achieve 80% accuracy on average by pixel-based quantitative performance evaluation that compares the automated segmentation outputs with the manually marked ground truth data. The presented technique greatly reduces labor-intensive workloads for pathologists, highly speeds up the process of TMA construction and allows further exploration of fully automated TMA analysis.
Ching-Wei Wang

Generalized Sparse Classifiers for Decoding Cognitive States in fMRI

The high dimensionality of functional magnetic resonance imaging (fMRI) data presents major challenges to fMRI pattern classification. Directly applying standard classifiers often results in overfitting, which limits the generalizability of the results. In this paper, we propose a new group of classifiers, “Generalized Sparse Classifiers” (GSC), to alleviate this overfitting problem. GSC draws upon the recognition that numerous standard classifiers can be reformulated under a regression framework, which enables state-of-the-art regularization techniques, e.g. elastic net, to be directly employed. Building on this regularized regression framework, we exploit an extension of elastic net that permits general properties, such as spatial smoothness, to be integrated. GSC thus facilitates simultaneous sparse feature selection and classification, while providing greater flexibility in the choice of penalties. We validate on real fMRI data and demonstrate how explicitly modeling spatial correlations inherent in brain activity using GSC can provide superior predictive performance and interpretability over standard classifiers.
Bernard Ng, Arash Vahdat, Ghassan Hamarneh, Rafeef Abugharbieh

Manifold Learning for Biomarker Discovery in MR Imaging

We propose a framework for the extraction of biomarkers from low-dimensional manifolds representing inter- and intra-subject brain variation in MR image data. The coordinates of each image in such a low-dimensional space captures information about structural shape and appearance and, when a phenotype exists, about the subject’s clinical state. A key contribution is that we propose a method for incorporating longitudinal image information in the learned manifold. In particular, we compare simultaneously embedding baseline and follow-up scans into a single manifold with the combination of separate manifold representations for inter-subject and intra-subject variation. We apply the proposed methods to 362 subjects enrolled in the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and classify healthy controls, subjects with Alzheimer’s disease (AD) and subjects with mild cognitive impairment (MCI). Learning manifolds based on both the appearance and temporal change of the hippocampus, leads to correct classification rates comparable with those provided by state-of-the-art automatic segmentation estimates of hippocampal volume and atrophy. The biomarkers identified with the proposed method are data-driven and represent a potential alternative to a-priori defined biomarkers derived from manual or automated segmentations.
Robin Wolz, Paul Aljabar, Joseph V. Hajnal, Daniel Rueckert

Optimal Live Cell Tracking for Cell Cycle Study Using Time-Lapse Fluorescent Microscopy Images

Cell cycle study using time-lapse fluorescent microscopy images is important for understanding the mechanisms of cell division and screening of anti-cancer drugs. Cell tracking is necessary for quantifying cell behaviors. However, the complex behaviors and similarity of individual cells in a dense population make the cell population tracking challenging. To deal with these challenges, we propose a novel tracking algorithm, in which the local neighboring information is introduced to distinguish the nearby cells with similar morphology, and the Interacting Multiple Model (IMM) filter is employed to compensate for cell migrations. Based on a similarity metric, integrating the local neighboring information, migration prediction, shape and intensity, the integer programming is used to achieve the most stable association between cells in two consecutive frames. We evaluated the proposed method on the high content screening assays of HeLa cancer cell populations, and achieved 92% average tracking accuracy.
Fuhai Li, Xiaobo Zhou, Stephen T. C. Wong

Fully Automatic Joint Segmentation for Computer-Aided Diagnosis and Planning

Orthopaedic examinations are a major reason for radiographic image acquisition. For many diagnostic problems measurements have to be computed from the recorded radiographs. As this task is time-consuming and lacks objectivity, it is desirable to perform these measurements automatically via so-called computational imaging. This requires robust and accurate methods trained and proven on clinical data.
We propose a fully automatic technique and present the three complementing stages of our segmentation algorithm. We evaluated the proposed method on more than 200 clinical images and achieve robust and precise delineations, well-suited for automated computation of orthopaedic measurements.
André Gooßen, Thomas Pralow, Rolf-Rainer Grigat

Accurate Identification of MCI Patients via Enriched White-Matter Connectivity Network

Mild cognitive impairment (MCI), often a prodromal phase of Alzheimer’s disease (AD), is frequently considered to be a good target for early diagnosis and therapeutic interventions of AD. Recent emergence of reliable network characterization techniques have made understanding neurological disorders at a whole brain connectivity level possible. Accordingly, we propose a network-based multivariate classification algorithm, using a collection of measures derived from white-matter (WM) connectivity networks, to accurately identify MCI patients from normal controls. An enriched description of WM connections, utilizing six physiological parameters, i.e., fiber penetration count, fractional anisotropy (FA), mean diffusivity (MD), and principal diffusivities (λ 1, λ 2, λ 3), results in six connectivity networks for each subject to account for the connection topology and the biophysical properties of the connections. Upon parcellating the brain into 90 regions-of-interest (ROIs), the average statistics of each ROI in relation to the remaining ROIs are extracted as features for classification. These features are then sieved to select the most discriminant subset of features for building an MCI classifier via support vector machines (SVMs). Cross-validation results indicate better diagnostic power of the proposed enriched WM connection description than simple description with any single physiological parameter.
Chong-Yaw Wee, Pew-Thian Yap, Jeffery N. Brownyke, Guy G. Potter, David C. Steffens, Kathleen Welsh-Bohmer, Lihong Wang, Dinggang Shen

Feature Extraction for fMRI-Based Human Brain Activity Recognition

Mitchell et al. [9] demonstrated that support vector machines (SVM) are effective to classify the cognitive state of a human subject based on fRMI images observed over a single time interval. However, the direct use of classifiers on active voxels veils the understanding of brain activity recognition at the neurological level. In this paper, we present neurological insights to this problem by introducing the covariance selection (CS) to model the correlations between active voxels. In particular, we show that new features, i.e., different correlations between active voxels, are valuable to the recognition of brain activities, in a sense that the fMRI image sequences of different brain activities exhibit quite dissimilar patterns after projection onto these features. Based on the new feature, we employ classifiers, e.g., SVM and nearest neighbor classifier, for brain activity recognition. Significant improvements are achieved compared against the method used in [9].
Wei Bian, Jun Li, Dacheng Tao

Sparse Spatio-temporal Inference of Electromagnetic Brain Sources

The electromagnetic brain activity measured via MEG (or EEG) can be interpreted as arising from a collection of current dipoles or sources located throughout the cortex. Because the number of candidate locations for these sources is much larger than the number of sensors, source reconstruction involves solving an inverse problem that is severely underdetermined. Bayesian graphical models provide a powerful means of incorporating prior assumptions that narrow the solution space and lead to tractable posterior distributions over the unknown sources given the observed data. In particular, this paper develops a hierarchical, spatio-temporal Bayesian model that accommodates the principled computation of sparse spatial and smooth temporal M/EEG source reconstructions consistent with neurophysiological assumptions in a variety of event-related imaging paradigms. The underlying methodology relies on the notion of automatic relevance determination (ARD) to express the unknown sources via a small collection of spatio-temporal basis functions. Experiments with several data sets provide evidence that the proposed model leads to improved source estimates. The underlying methodology is also well-suited for estimation problems that arise from other brain imaging modalities such as functional or diffusion weighted MRI.
Carsten Stahlhut, Hagai T. Attias, David Wipf, Lars K. Hansen, Srikantan S. Nagarajan

Optimal Gaussian Mixture Models of Tissue Intensities in Brain MRI of Patients with Multiple-Sclerosis

Brain tissue segmentation is important in studying markers in human brain Magnetic Resonance Images (MRI) of patients with diseases such as Multiple Sclerosis (MS). Parametric segmentation approaches typically assume unimodal Gaussian distributions on MRI intensities of individual tissue classes, even in applications on multi-spectral images. However, this assumption has not been rigorously verified especially in the context of MS. In this work, we evaluate the local MRI intensities of both healthy and diseased brain tissues of 21 multi-spectral MRIs (63 volumes in total) of MS patients for adherence to this assumption. We show that the tissue intensities are not uniform across the brain and vary across (anatomical) regions of the brain. Consequently, we show that Gaussian mixtures can better model the multi-spectral intensities. We utilize an Expectation Maximization (EM) based approach to learn the models along with a symmetric Jeffreys divergence criterion to study differences in intensity distributions. The effects of these findings are also empirically verified on automatic segmentation of brains with MS.
Yiming Xiao, Mohak Shah, Simon Francis, Douglas L. Arnold, Tal Arbel, D. Louis Collins

Preliminary Study on Appearance-Based Detection of Anatomical Point Landmarks in Body Trunk CT Images

Anatomical point landmarks as most primitive anatomical knowledge are useful for medical image understanding. In this study, we propose a detection method for anatomical point landmark based on appearance models, which include gray-level statistical variations at point landmarks and their surrounding area. The models are built based on results of Principal Component Analysis (PCA) of sample data sets. In addition, we employed generative learning method by transforming ROI of sample data. In this study, we evaluated our method with 24 data sets of body trunk CT images and obtained 95.8 ± 7.3 % of the average sensitivity in 28 landmarks.
Mitsutaka Nemoto, Yukihiro Nomura, Shohei Hanaoka, Yoshitaka Masutani, Takeharu Yoshikawa, Naoto Hayashi, Naoki Yoshioka, Kuni Ohtomo

Principal-Component Massive-Training Machine-Learning Regression for False-Positive Reduction in Computer-Aided Detection of Polyps in CT Colonography

A massive-training artificial neural network (MTANN) has been investigated for reduction of false positives (FPs) in computer-aided detection (CAD) of lesions in medical images. The MTANN is trained with a massive number of subvolumes extracted from input volumes; hence the term “massive training”. A major limitation of this technique is a long training time due to the high input dimensionality. To solve this problem, we incorporated principal-component (PC) analysis for dimension reduction into the MTANN framework, which we call a PC-MTANN. To test the PC-MTANN, we compared it with the original MTANN in FP reduction in CAD of polyps in CT colonography. With the use of the dimension reduction architecture, the time required for training was reduced substantially from 38 to 4 hours, while the original performance was maintained, i.e., a 96% sensitivity at an FP rate of 3.2 and 3.0 per patient by the original MTANN and the PC-MTANN, respectively.
Kenji Suzuki, Jianwu Xu, Jun Zhang, Ivan Sheu


Weitere Informationen

Premium Partner