Abstract
In the recent years, applications of computer vision system on the biological and medical images are increasing. In clinics, the computer vision system is helpful for visual interpretation. Among the diseases, the liver disease has received much attention. We attempt to segment out the region which is likely to be ill from the ultrasonic liver images. Furthermore, we may identify the normal region and suspected ill region. At first, we extract the feature vector from the ultrasonic liver images, and then the feature vector is applied into the clustering algorithm. The feature extraction algorithm is based on the spatial-frequency decomposition and fractal geometry. Such a multiresolution feature vector has been proved trustworthy when handling such texture images as ultrasonic images. In the thesis, we compare the performance of the two decomposition methods: (1) M-band wavelet transform and (2) Gabor Transform. Besides, we also analyze the performance and computational complexity versus different feature lengths. We attempt to reduce the feature length to improve the execution speed but not to influence the performance too much as the premise. After getting efficient feature vector, the clustering algorithm is applied. Many clustering algorithms have been proposed in the past. K-means and Fuzzy C-means (FCM) have great performance in most images, and Self Organizing Maps (SOM), which can describe the structure of feature dimension in the neural network filed, is also applied to many clustering cases. In the thesis, the performances of the above three clustering methods combined with the two decomposition methods are compared. An efficient initialization procedure is proposed to solve the main shortcoming of K-means algorithm and to improve the performances of other clustering methods. At last, a modified FCM method is proposed, which provides an elastic tuning operation based on the probability maps to adjust the segmentation result. A great performance is obtained in the cases of nature texture images via the mechanism of this tuning operation. The tests are extended to the cases of ultrasonic liver images. The experiments show that the unsupervised segmentation framework can efficiently segment out the rough area concerned and the modified FCM we proposed is furthermore expected to provide doctors a probability tuning operation according to clinical experience. Instead of hard-clustering, the modified FCM outperforms K-means, FCM and SOM by its elasticity. A quantitative characterization based on the proposed unsupervised segmentation algorithm not only can be utilized to establish an automatic computer-aided diagnostic system. As well, to increase the visual interpretation capacity of ultrasonic liver image for junior physicians, an off-line learning system can be developed to investigate the visual criteria.