Computer-Aided Diagnosis Of Liver Space-occupying Lesions Based On Ultrasonic Medical Image Analysis

The liver cancer, a hazard to people's health, is common in China, with the mortality second to gastric cancer. The population dying of liver cancer in China accounts for 42% of that in the world. With its unique advantages as no radiation, real time, repeatability, low cost and expedience to operate, etc., the ultrasound diagnosis has become an effective measure for diagnosing the space-occupying liver lesions. The ultrasound clinical diagnosis, however, relies on doctors' unaided viewing and estimating, thus the diagnosis results heavily depend on diagnosis physicians' clinical experience, resulting in probably misdiagnosis to the nature of the space-occupying lesions. As a result, the malignant cancer sufferers may miss the optimum period for treatment, which has a strong impact on recovery. An objective, timely and accurate computer-aided diagnosis is urgently necessary.The research on the recognition of liver space-occupying lesions in ultrasound images is still at its early stage currently in the world. In this paper, the recognition of liver space-occupying lesions in ultrasound images is studied and developed, with the results as a computer-aided diagnosis means for ultrasound diagnosis of liver space-occupying lesions.Experiments are done on a total of 280 cases of liver images, including 112 cases of normal liver images, 90 cases of liver cancer images, 38 cases of liver hemangioma images and 40 cases of liver cyst images. From the standpoint of the criterion by which clinicians diagnose liver space-occupying lesions, the clinical diagnosis criterion is abstracted into numeric features, and the recognition of liver space-occupying lesions in ultrasound images is studied by using ultrasound image texture analysis and pattern recognition techniques. This study contains the following sections: 1. Acquire the ultrasound images of normal liver, liver hemangioma, liver cyst and liver cancer of various kinds and stages, and make conversion to image formats as pre-processing; 2. In accordance with clinical diagnosis criterion, abstract two selection rules of ROI, and extract manually the ROI in each image under the guidance from doctors of rich experience; 3. According to the selected ROI, extract relevant features in each image; 4. Acknowledge the supervised recognition method: back-propagation (BP) artificial neural network (ANN) algorithm structure -three-level BP neural network. According to the network structure and the function of each level, select the features by jointly using U test as well as feature selection methods based on correlation analysis and on quadratic mutual information, and obtain the suboptimum feature vectors of each level network; 5. Determine the dimension of the suboptimum feature vectors of each level network through experiments, and train & test independently each level with the leave-one-out training & testing method; then cascade each level network to generate the computer-aided diagnosis system, and carry out recognition at different stages to four modalities of liver images as normal liver, liver cancer, liver hemangioma and liver cyst.The experimental results show that: 1. The feature vector comprising 8 features is capable to recognize normal livers from abnormal ones, including mean-intensity ratio between tumor and normal, mean convergence ratio, mean-intensity ratio between normal and halo, contrast ratio, variance ratio, mean-intensity ratio between halo and tumor, entropy ratio and code entropy ratio; 2. The feature vector comprising 5 features is capable to recognize liver cysts from abnormal livers, including mean-intensity ratio between tumor and normal, sum entropy ratio, mean convergence ratio, entropy ratio and mean-intensity ratio between halo and tumor; 3. The feature vector comprising 8 features is capable to recognize liver cancers and liver hemangioma, including circularity 2, area ratio, variance ratio, mean-intensity ratio between normal and halo, mean convergence ratio, entropy ratio, mean-intensity ratio between tumor and normal and entropy of the normalized radial;4.The proposed BP ANN is promising in recognition of liver space-occupying lesions. The precise recognition rates of liver cancer, liver hemangioma, liver cyst and normal liver are 100%, 94.7%, 95% and 100%, respectively. The findings of this study can serve as a computer-aided diagnosis means for clinically diagnosing liver cancer, liver hemangioma and liver cyst.