11th International Conference on ICTIS

Authors - Diego Perez-Lopez, Rodolfo Bojorque, Jorge Duenas-Lerin, Raul Lara-Cabrera
Abstract - Accurate early detection of liver cancer remains a significant clinical challenge, primarily due to scarce annotated imaging data, inconsistencies in radiological interpretation, and the inherent opacity of deep learning models. To address these limitations, this study proposes a clinically informed, explainable deep learning framework designed specifically for low-annotation settings. The framework combines transfer learning with advanced visualization techniques, enabling both high diagnostic accuracy and medically meaningful outputs that integrate seamlessly into clinical workflows. Three pre-trained CNN architectures — ResNet-50, DenseNet-121, and EfficientNet-B4 — were adapted to liver cancer imaging through domain-specific fine-tuning. Model generalizability was reinforced by combining geometric data transformations with StyleGAN2-derived synthetic lesion generation. Model transparency was facilitated through Gradient-weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP), while clinical trustworthiness was evaluated via predictive uncertainty quantification, subgroup bias analysis, and resistance to adversarial perturbations. The proposed framework was evaluated on the LiTS and TCGA-LIHC datasets, demonstrating a 15–20% improvement in accuracy over baseline models that consisted of standard convolutional neural networks trained from scratch without transfer learning or data augmentation. EfficientNet-B4 achieved 94.2% accuracy, 0.96 specificity, and an AUC-ROC of 0.978. Grad-CAM accurately highlighted tumor regions in 89.4% of cases, and Bayesian dropout identified 7.3% of predictions as uncertain. These findings demonstrate the framework’s potential for clinical deployment by balancing performance, transparency, and reliability.