Machine Learning Approaches To Medical Image Analysis

This thesis presents a study of the Computer Aided Detection (CAD) tasks for automated medical image analysis and proposes two new methods to realize the CAD systems. Based on the proposed methods, this thesis constructs the integrated system for detection of the lesions in medical images. Using the digital mammograms, this thesis designs, validates, and analyzes the proposed methods and the integrated detection system.This thesis regards the image analysis process of CAD as two stages. For the first stage, including the steps of image conditioning and image segmentation, this thesis proposes the adaptive image segmentation method using the reinforcement learning techniques, which selects the processing algorithms and adjusts the parameter settings appropriately so as to attain the optimal results. For the second stage, including the steps of feature extraction and object recognition, this thesis proposes the automated lesion detection method using the representation of multiresolution histogram features and the kernel classification algorithms, which eliminates or restricts the step of feature selection so as to detect various kinds of lesions simultaneously, and as a result, it is unnecessary to detect each kind of lesions respectively. Combining the proposed two methods, this thesis constructs the integrated detection system and optimizes the performance by appropriate adjustments.The experimental results show that: the adaptive image segmentation method can obtain high segmentation precision for various kinds of images, the automated lesion...