Classification of Focal Liver Lesions Using Deep Learning with Fine-Tuning

Liver cancer is one of the leading causes of death worldwide. Computer-aided diagnoses play an important role in liver lesion diagnoses (classification). Recently, several deep-learning-based computer-aided diagnosis systems have been proposed for the classification of liver lesions. The effectiveness of these systems has been demonstrated; however, the main challenge in deep-learning-based medical image classification is the lack of annotated training samples. In this paper, we demonstrate that transfer learning and fine-tuning can significantly improve the accuracy of liver lesion classification, especially for small training samples. We used the residual convolutional neural network (ResNet), which is a state-of-the-art network, as our baseline network for focal liver lesion classification using multi-phase CT images. Fine-tuning significantly improved the classification accuracy from 83.7% to 91.2%. This classification accuracy (91.2%) is higher than that of state-of-the-art methods.