Liver Ultrasound Image Classification of Periductal Fibrosis Based on Transfer Learning and FCNet for Liver Ultrasound Images Analysis System

The current population of Southeast Asia is found to have died of Cholangiocarcinoma (CCA) approximately 28,000 each year. CCA risk factors, particularly the finding of Periductal Fibrosis (PDF), were observed and measured by analyzing the ultrasonography (US) image. The CCA Screening and Care Program (CASCAP) carry and store the enormous US images in their data warehouse server to facilitate the online diagnosis by the experts. While the amount of data increasing but the expert whose responsibility to determine the existence of PDF in US image is less dramatic. This leads to a decrease in the survival rate of the patients. Due to this crisis problem, we proposed a structure of transfer learning to classify the stages of PDF which is being used in the development of a Liver Ultrasound Image Analysis System (LUIAS). We also introduced data augmentation to boost the characteristic of the PDF criterion in the data preparation step, which is designed by modification of the Discrete Fourier Transform. Cross-validation is applied to the learning process to evaluate the overall performance of the structure. The experimental results show the proposed method can reach a higher accuracy compare to the conventional technique, which is 0.92 and 0.81 respectively. The best stage of the learning process is also deep copied to be further used as a classifier in practical operation at LUIAS.