The Role of Convolutional Neural Networks in the Automatic Recognition of the Hepatocellular Carcinoma, Based on Ultrasound Images

Hepatocellular carcinoma (HCC) represents the most frequent form of liver cancer. It evolves from cirrhosis, as the result of a restructuring phase at the end of which dysplastic nodules appear. HCC is the main cause of death in people affected by cirrhosis. The most reliable method for HCC diagnosis, the golden standard, is the needle biopsy, but this is an invasive technique, dangerous for the human body. We develop computerized methods, based on ultrasound images, in order to perform automatic diagnosis of HCC in a noninvasive manner. In our previous research, we elaborated the textural model of HCC, based on classical and advanced texture analysis methods, in combination with traditional classification techniques, which led to a satisfying accuracy. We aim to improve this performance in our current and further research. In this article, we analyzed the role of specific deep learning techniques concerning the automatic recognition of HCC from ultrasound images. We chose the Convolutional Neural Networks (CNN), this method being well known for its performance in the field of image recognition. Thus, CNN are based on artificial neural networks, they also performing image processing operations, such as inner convolutions. In order to evaluate the newly adopted technique, we assessed the classification accuracy achieved in the cases of distinguishing HCC from the cirrhotic parenchyma on which it had evolved, respectively when differentiating HCC from the hemangioma benign liver tumor. We compared the accuracy of CNN with our previous results, based on texture analysis methods and it resulted that CNN yielded better recognition rates than the classical texture analysis techniques, respectively comparable with those provided by the advanced texture analysis methods. However, further improvements will be necessary, which we mention within the last section of this article.