Learning-based Curvilinear Structure Analysis in Medical Images

Analysis of curvilinear structures is an important problem in computer-aided diagnosis and image guided interventions with applications such as vessel structure classification and branching structure detection in mammographic images, vessel segmentation in retinal images, deformable tracking of curvilinear structures in X-ray sequences. Curvilinear structure analysis is often challenging due to large variations in their appearances and profiles, as well as image noise. Also, low visibility and poor image quality due to a low dose of radiations in interventional imaging make the task more difficult, especially in breast mammographic images and dynamic X-ray images. Furthermore, for curvilinear structure tracking, these structures undergo complex deformable motion as a result of complex 3D anatomical movements projected onto the 2D image plane. In this dissertation, we mainly focus on curvilinear structure analysis of mammographic images, retina images and X-ray images. These tasks include vessel structure classification, branching structure detection, vessel segmentation and curvilinear structure deformable tracking.;For vessel structure classification, we present a framework using the bag-of-words model and histogram intersection (HI) similarity measure. We first use the bag-of-words model for image representation, which captures the texture information by collecting local patch statistics. Then, we propose applying normalized histogram intersection as similarity measure. Finally, the classification is achieved by combining KNN classifier or support vector machines (SVM). The proposed method is evaluated on a galactographic dataset and compared with several previously used methods. We show that both normalized HI and HI+SVM outperform previous state-of-the-art methods. We also notice that, when using KNN classifiers, normalization is an important step that helps to improve the accuracy of histogram similarities.;For branching structure detection in mammographic images, we describe an approach to automatically detect branching structure region-of-interest (ROI) in clinical breast images. We develop a boosting-based framework using AdaBoost algorithm and Haar wavelet features. In order to keep a single ROI of an input image in the detection, candidates ROIs with spatial overlap are merged according to their confidence scores. We compare three filtering strategies to eliminate false positives. These strategies differ in the approach to fuse confidence scores by summation, averaging or selecting the maximum one. Experiments on clinical galactograms show the presented false positive filtering strategies achieve promising results.;For curvilinear structure segmentation, we present a learning-based framework for vessel segmentation on mammographic images and retinal images. Typically, our proposed framework is composed by ensemble learning algorithm and hybrid features to represent an instance. In vessel segmentation of mammographic images, the ensemble learning algorithm is a forest with boosting trees. The feature pool contains local, Gabor and Haar features. Our method is tested on a real dataset with 20 anonymous mammographic images and achieves 10.06% equal error rate of breast vessel segmentation. Similarly, we extend the work on vessel segmentation in breast mammographic images to retinal images. We design a hybrid feature pool containing recently invented descriptors including the stroke width transform (SWT) and Weber's local descriptors (WLD), as well as classical local features including intensity values, Gabor responses and vesselness measurements. Secondly, we encode context information by sampling the hybrid features from an orientation invariant local context. The ensemble learning algorithm is a random forest to fuse the rich information encoded in the hybrid context-aware features. We apply the proposed method to retinal vessel segmentation and evaluate it using three publicly available datasets: the DRIVE dataset, the STARE dataset and the High-Resolution Fundus (HRF) Image Database (HRFID). Quantitative evaluation results demonstrate the effectiveness of our approach.;For curvilinear structure deformable tracking, we introduce two methods to address the problem of robust tracking of vascular structure and intravascular devices in X-ray images. The first approach uses the Bayesian filtering framework and data driven measurement models to estimate the deformable motion field. We first convert the maximum likelihood estimation of the motion field to an energy minimization problem, and then use a variational solution to solve it. A randomized regression forest is employed to learn the probability density function of the measurements from training samples with known displacements. The results demonstrate that our approach outperforms the registration-based solution. The second approach uses the tensor-based algorithm with model propagation. Specifically, the deformable tracking is formulated as a multi-dimensional assignment problem which is solved by rank-1 ℓ1 tensor approximation. The model prior is propagated in the course of deformable tracking. Both the higher order information and the model prior provide powerful discriminative cues for reducing ambiguity arising from the complex background, and consequently improve the tracking robustness. The results show, both quantitatively and qualitatively, that our approach achieves a mean tracking error of 1.4 pixels for vascular structure tracking and 1.3 pixels for catheter tracking.