Research Of Computer-Aided Fatty Liver Diagnosis Based On Ultrasonic Medical Image Analysis

With the improvement of living conditions, fatty liver is now being diagnosed more frequently. It will cause abnormal liver tests, inflammation and possibly permanent liver damage when left untreated. With the fast development of medical ultrasound, it's now possible to diagnose fatty liver by B- scan ultrasound. But the diagnosis is strongly dependent on the experiences of physicians, so the result may be subjective and not very reliable. Some related research works indicate that several image features, e.g. the second order statistics extracted from the Gray Level Co-occurrence Matrices (GLCM) and the Gray Level Run-Length (GLRL) matrices, can be used to distinguish between the normal liver and the fatty liver. But most of the research works were based on the direct comparison of the features or the hypothesis tests of two groups of samples, and there are not any fatty liver recognizing techniques available yet. The objective of this thesis is to provide a computer-aided method for the diagnosis of fatty liver by B-scan ultrasonic imaging. There are not any related reports yet.This thesis focuses on the researching and developing works of the fatty liver diagnosing techniques based on image analysis and image recognition. Firstly, B-scan ultrasound images are captured form patients and the regions of interest are localized. During the feature extraction approach, three models are employed including mean intensity ratio of the near field and the far field, the gray level co-occurrence matrices and the gray level run-length matrices. 10 statistics are extracted from the three models for each image. After the feature selection approach which involves hypothesis tests and artificial neural networks, the best feature vectors are created. Finally, C-means clustering algorithm, Self-Organized Feature Mapping (SOFM) artificial neural network and Back-Propagation (BP) Artificial Neural Network are respectively employed to classify the sample vectors. There are totally 44 cases used during this work. 33 of them are normal liver cases, and the other 11 ones are fatty liver cases. After the feature selection approach, there are only 4 features left for further researches including the Angular Second Moment (ASM), Entropy and Inverse Differential Moment from the GLCMs, as well as the Mean Intensity Ratio. Thus, feature vectors which indicated two classes of images are created with the four features. The accuracy rates of C-means clustering algorithm are 100% for normal liver samples and 63.6% for fatty liver samples. The results of the SOFM neural network show that the accuracy rates are 93.94% for normal liver samples and 100% for fatty liver samples. The accuracy rates of BP neural network are 100% for normal liver samples and 100% for fatty liver samples. From the results list above, it can be concluded that C-means algorithm's performance is not very good, especially for the recognition of fatty liver. This is because the distribution overlaps of the samples and the essential shortcoming of the algorithm itself. In contrast to those of C-means algorithm, the results of SOFM neural network are much better. But there are several normal liver samples that can not be recognized by SOFM network. The supervised approach - BP neural network has the best performance, because not only the samples themselves but also their class information is both used in the approach.