Abstract
Background: Hepatocellular carcinoma (HCC) is a significant global health concern that requires early detection for effective treatment. Objectives: The objective of this study was to develop a system for screening HCC in B-mode ultrasound images. Materials and methods: The dataset consisted of 1665 hemangioma (HEM) images, including 961 typical HEM, 704 atypical HEM, and 543 HCC images. Four YOLOv4 models were trained: one for HCC detection, one for the conventional two-class detection of HEM and HCC, one to detect typical HEM and suspicious lesions, and the last one was our two-stage model consisting of a detector and classifier. In the first stage, a YOLOv4-based detector with ResNet-50 as the backbone was used to identify focal liver lesions. The second stage utilized ResNet-50 as a classifier to classify the lesions into HCC, atypical HEM, or typical HEM. Differentiating between HCC and atypical HEM is not necessary, as both require further investigation with CT or MR imaging. Results: The evaluation of the developed HCC screening system using ten-fold cross-validation showed that grouping HCC and atypical HEM together significantly increased precision from 0.74 to 0.88 and improved HCC recall from 0.64 to 0.68. Furthermore, employing the two-stage method further improved HCC recall from 0.68 to 0.72. Conclusion: The results indicate that combining HCC and atypical HEM into a single class and using a two-stage approach for detection led to substantial improvements in precision and HCC recall. These findings highlight the potential of the developed system for effective HCC screening in B-mode ultrasound images. The two-stage method provided better detection than the detector-only method. More accurate detection was achieved when lesions were classified based on appearance and clinical protocols.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have