Knowledge distillation–enabled ResNet framework for hepatocellular carcinoma diagnosis on multiphase CT: Toward scalable global deployment.

Abstract

4156 Background: Hepatocellular carcinoma (HCC) accounts for 85–90% of primary liver cancers and remains a leading cause of cancer-related mortality worldwide. Early and accurate imaging-based diagnosis is critical for staging, treatment selection, and survival. Multiphase contrast-enhanced CT is widely used for HCC detection, yet diagnostic accuracy varies due to lesion heterogeneity and radiologist workload. Deep learning models demonstrate high performance but are computationally intensive, limiting deployment in resource-constrained settings. Knowledge distillation enables transfer of diagnostic capability from large teacher models to lightweight student networks while preserving accuracy. We evaluated a ResNet-based distillation framework for efficient HCC diagnosis using multiphase CT imaging. Methods: We analyzed a curated dataset of 140 pathologically confirmed HCC cases, comprising arterial, portal venous, and delayed-phase CT images annotated by two expert radiologists independently . A high-capacity ResNet152 teacher model was trained to detect and localize HCC lesions using hierarchical residual feature learning across multiphase inputs. A compact ResNet18 student model was trained via temperature-scaled knowledge distillation using soft probability targets and regularized cross-entropy loss. Performance was evaluated using accuracy, sensitivity, specificity, F1 score, and AUROC. Computational efficiency and inference latency were assessed to determine real-world deployment feasibility. Results: The ResNet152 teacher achieved robust diagnostic performance ( > 97% accuracy) for HCC detection across lesion sizes and locations. The distilled ResNet18 preserved near-teacher accuracy, achieving AUROC exceeding 0.93 with balanced sensitivity and specificity ( > 93%). Knowledge distillation reduced model parameters by over 80% and substantially lowered inference time, enabling real-time execution on standard clinical hardware. Performance remained consistent across lesion diameter and hepatic segment subgroups, comparable to expert radiologist assessment Conclusions: Knowledge distillation enables a lightweight ResNet18 model to retain near–ResNet152 diagnostic performance for HCC detection on multiphase CT while markedly reducing computational requirements. This scalable framework addresses key barriers to AI adoption in oncology imaging and supports global deployment of AI-assisted HCC diagnosis. Prospective multicenter validation is warranted to assess clinical impact on early detection and treatment planning.