Automatic Detection And Classification Of Focal Liver Lesion In Ultrasound Images Based On Sparse Representation

China is the high-prone area of liver cancer. The population of liver cancer and death in China accounts for half of that in the world. Liver cancer has already been a serious threat to people's health. Research has shown that, early detection and accurate diagnosis for liver cancer can reduce the mortality effectively. With its non-radiation, easy to operate, inexpensive and non-invasive, the ultrasound diagnosis has become the main measure for diagnosing liver diseases on clinical practice.Due to the defects inherent in the ultrasound image and the complexity appearance of diseases, it's a challenging task for liver diseases diagnosis in ultrasound images. What's more, doctors have to face enormous of work every day leading to a higher probability of misdiagnosis.With the development of computer technology and medical image, computer-aided diagnosis(CAD) has shown immeasurable effect on automatic classification for liver diseases in ultrasound images. In this paper, we focus on the automatic detection and diagnosis of focal liver lesion in ultrasound images aiming to provide effective reference for doctors to diagnose and reduce the workload of doctors.The contribution of this paper can be summarized as follows:1. For detecting lesion region more precisely, we propose an automatic ultrasound images segmentation based on multi-dictionaries sparse representation and energy minimizing. The proposed method can automatically select seed point as the starting of region growing with multi-dictionaries sparse representation. We use graph theory to obtain an energy function which represents the structure of lesion region and the boundary information. Then the region growing process is controlled by the energy function. The experiment results have shown that our proposed method can delineate the lesion region effectively and accurately.2. According to the pathological characteristics of lesions and their diversity appearances in ultrasound images, we design a specific image feature. Considering the varies of ultrasound image quality caused by the difference of ultrasound devices and divergence between individual organizations, we not only extract the texture feature of lesion region and also take normal region as reference. We analysis both inside and outside of lesion region to improve reliability and robustness. What's more, considering the traditional features have poor description to the pathology and ultrasonic appearance, we specially design a new image feature which can represent the complexity of the lesion inner structure effectively and be discriminative to liver lesions.3. To achieve the precise classification of focal liver lesions, we use SRC classification framework. Aiming to solve the problem in dictionary construction of SRC, we propose a method for dictionary expansion based on sparse reconstruction to construct a good dictionary. The dictionary constructed by our method is redundant and has a compact inner structure, which leads to a better performance in classification. Experimental results have shown that the proposed method could improve the classification accuracy in comparison with other state-of-the-art classifiers. It should be capable of assisting the physicians for liver disease diagnosis in the clinical practice.