Transfer Learning versus Custom CNN Architectures in NAFLD Biopsy Images

Nonalcoholic fatty liver disease (NAFLD) is one of the most frequent liver conditions representing a wide range of intrahepatic disorders, varying from steatosis to nonalcoholic steatohepatitis (NASH). Steatosis refers to the accumulation of benign fat cells, which at higher rates leads to NASH progression, as the major risk factor for hepatic fibrosis and cirrhosis, as well as for hepatocellular carcinoma (HCC). In recent years the medical field has focused on preventing the progression of these diseases, with microscopic biopsy images being the gold standard imaging modality in modern clinical trials. The proposed work aims at the high classification ability of four histological liver structures, by training a convolutional neural network (CNN) and comparing its diagnostic performance with various pre-trained deep CNN architectures. All diagnostic attempts were made on an augmented image dataset, with the new CNN model achieving a 95.8% classification accuracy, while AlexNet emerging as the most efficient architecture with a corresponding performance of 97.8%.