Hepatic Lesion Recognition Based on Deep Visual Feature Learning

There are many kinds of hepatic lesions which threaten human healthy by high incidence. Analyzing CT images is an effective method to evaluate the type of hepatic lesions. The original evaluation method is that medical personnel assessment by images, but abdominal CT images are complex, and the accuracy of the discrimination depends strongly on the subjective factors, knowledge level and experience of medical personnel. With the development of artificial intelligence technology, the hepatic lesions recognition technology based on deep visual feature learning can achieve a higher accuracy. While reducing the workload for doctors, this technology can assist doctors to diagnose the disease, so as to obtain more objective and accurate diagnosis results. The database established in this paper mainly includes three kinds of CT images: primary liver cancer, cirrhosis and normal liver. After the database pretreatment with image de-noise algorithms based on median filter and picture enhancement algorithms based on histogram equalization, the convolution neural network CaffeNet was taken as the main framework, and a large number of labeled liver CT images were used to train the network in the experimental process. Meanwhile, the network parameters were continuously updated according to the cost function. Traditional hepatic lesion recognition technology only depends on the surface features from the image, such as gray scale, statistical structure and texture. Deep learning, a tool with high precision and robustness, could be used to mine the hidden features in the huge database. Through testing the trained deep learning model, the average recognition accuracy of 96.67% and the average single sample recognition time of 1.28 seconds could be achieved perfectly. Therefore, there is no doubt that hepatic lesion recognition technology based on deep visual feature learning will greatly liberate doctors' physical strength and promote the development of pathology in the near future.