Binary Classification of Benign and Malignant Hepatic Lesions with Portal Venous Phase Computed Tomography Images with Deep Learning: A Single-institution Study

Abstract Background: Differentiate benign and malignant hepatic lesions with a limited phase of CT scan is a practical clinical challenge. Aim: To evaluate the feasibility of differentiation of benign and malignant hepatic lesions with portal venous phase computed tomography (CT) images with deep learning. Methods: This was a retrospective single-center study; patients with focal hepatic lesions were detected in the contrast-enhanced CT scan from January 2016 to December 2018 were enrolled. Benign hepatic lesions enrolled in this study include hepatic cysts and hepatic hemangioma; malignant lesions include hepatocellular carcinoma (HCC) and hepatic metastases. Portal venous phase CT images of focal hepatic lesions were labeled by one radiologist with 6-year experience in abdominal radiology. Labeled CT images were input into convolutional neural network (CNN) to perform binary classification with training-test split of 70%:30%. The performance was evaluated by precision, recall, F1-score, and receiver operating characteristic curves. Results: There were 214 images of benign hepatic lesions and 223 images of malignant hepatic lesions enrolled in this study, including 84 hemangiomas, 131 hepatic cysts, 146 HCC, and 77 hepatic metastatic lesions. CNN model achieved the best results to classify benign and malignant hepatic lesions, with an average test precision of 0.91, recall of 0.91, and F1-score of 0.91. The area under the curve was 0.96. Conclusion: With portal venous phase CT images, CNN can perform binary classification of benign and malignant hepatic lesions with acceptable accuracy in a limited dataset.