Study Of Machine Learning Techniques And Applications In Med-ical Image Analysis

The rapid development of the clinical equipment greatly accelerates the study of medical image analysis. Among different kinds of methods, the machine learning-based methods play a significant role, and has already became one of the most widely used approaches. The thesis tries to apply our developed machine learning techniques to solve the clinical problems within the two applications:’lung cancer image identifi-cation based on needle biopsies specimens’ and ’prostate segmentation for CT image guided radiotherapy’. The details can be concluded as follows:(1) Transductive cost-sensitive learning:to address the problem that misclassifi-cation costs are naturally unequal among different classes in lung cancer image clas-sification (misclassifying the normal image as cancerous just require the pathologists for re-diagnosis, while misclassifying the cancerous image as normal means the patient might lose the best chance for survive), we introduce the cost-sensitive learning setting in the application. Moreover, to further alleviate the labeling burden for pathologists on large-scale samples, we introduce the transductive learning, aiming to achieve the best possible results with limited numbers of training images. Based on these observa-tions, we proposed a novel transductive cost-sensitive classifier:mCLRLS, which has demonstrated its effectiveness on the real lung cancer image set.(2) Multi-modal sparse representation-based classification:to further improve the classification results especially for lung cancer types classification, we proposed a nov-el methods, named’multi-modal sparse representation-based classification’(mSRC), to study the multi-modal information for classification. mSRC aims to use the large diversity criterion originally proposed in learning theory, to guarantee both the inter-modal discriminative ability and intra-model diversity. mSRC can be well solved by using genetic algorithm, and finally a hierarchical fusion strategy is designed for pre-diction. The experimental results can validate the usefulness of the assumptions.(3) Interactive prostate segmentation for CT image guided radiotherapy:tradi-tional learning-based methods for prostate segmentation are well designed for patients with relatively small prostate motion. To address the issue coming from large prostate motion, we try to develop a method, which only requires a little time from physi-cian, in order to largely increase the segmentation accuracy. Specifically, we propose a patient-specific ROI extraction method (IPRE), a transductive feature selection method (tLasso), as well as regression method (wLapRLS). The results on the real CT images validate the proposed method can obtain superior results compared with several related works.(4) Spatial-constrained transductive feature selection:for most learning-based prostate segmentation methods, feature selection plays an important role, but previ-ous methods belong to ’global feature selection’. To better address the local prostate appearance variation problem, we proposed a novel feature selection method Spatial-COnstrained Transductive LassO (SCOTO) for selecting the most informative features for individual local regions as well as keeping the smoothness for neighboring blocks. Moreover, a gradient descent-based strategy is designed for optimization. From the experimental analysis, proposed method outperforms several state-of-the-arts.