Explainable and Trustworthy AI for Liver Cancer Diagnosis Using Transfer Learning with Uncertainty, Fairness, and Robustness Evaluation

Abstract

Accurate diagnosis of liver cancer depends not only on predictive accuracy but also on a transparent account of
how a model arrives at its decisions. This study proposes a framework that explicitly targets interpretability and
reliability within a transfer-learning setting. A convolutional neural network pre-trained on large-scale data is
adapted to medical images, with the aim of mitigating limited annotation availability while preserving
informative representations. For interpretability, Grad-CAM is employed to localize image regions that most
strongly influence the model’s predictions, and SHAP is used to quantify the contribution of input features to
the output. Beyond explainability, the framework examines reliability through uncertainty estimation, fairness
analyses across predefined subgroups, and robustness evaluations under controlled input perturbations.
Experiments on publicly available datasets indicate that the proposed approach attains competitive diagnostic
performance while offering additional evidence about model behavior. Overall, the results support the view that
transfer learning, when paired with explanation methods and reliability assessment, may contribute to more
dependable AI-assisted diagnosis in clinical contexts.