Abstract: |
Accurate liver cancer classification is essential, as it substantially influences the selection of effective treatment strategies and impacts patient prognosis. Convolutional neural network (CNN) classifiers typically require extensive labeled datasets for training to attain decent performance. However, the process of obtaining labeled data through manual labeling is time-consuming, potentially biased, and costly when applied to large datasets. This study utilizes the Simple Siamese (SimSiam) contrastive self-supervised learning approach to enhance the classification of liver tumours, especially considering the limited availability of labeled computed tomography (CT) scans of liver cancer. We integrate SimSiam with three baseline CNN-based classifiers - Inception, Xception, and ResNet152 - and pretrain them with two loss functions: mean squared error (MSE) and cosine similarity (COS). Our findings show consistent improvements for three classifiers compared to the baseline models. Specifically, the ResNet152 model exhibits the highest performance among the evaluated networks. With MSE and COS losses, the classification accuracy for ResNet152 improves by 1.27% and 2.53%, respectively. The classification accuracy of the Inception model improves by 3.95% and 5.26%. Similarly, Xception’s validation accuracy demonstrates an increase of 2.60% with both loss functions, compared to the baseline models. We validate our pipeline via our multi-resolution in-house abdominal CT scans of primary and secondary liver cancers, including 155 patients with hepatocellular carcinoma, 198 patients with colorectal liver metastases, and 107 patients with intrahepatic cholangiocarcinoma. Source code available at:https://github.com/Ramtin-Mojtahedi/SimSiam-LiverCancer-CL. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023. |