A Unified Explainable AI Framework for Reliable Clinical Decision Support and Medical Image-Based Cancer Classification

Abstract

Artificial intelligence (AI) can match specialist accuracy in cancer detection, yet adoption in clinical practice remains limited because most high-performing models are opaque, poorly calibrated and trained on data that cannot be shared across institutions. This paper proposes a unified explainable AI (XAI) framework that integrates (i) a transfer-learning convolutional neural network for medical images, (ii) a calibrated soft-voting ensemble for tabular cytological and clinical features, (iii) complementary explanation methods (Grad-CAM for images and SHAP for features), (iv) a confidence-based uncertainty gate that refers ambiguous cases to clinicians, and (v) a federated privacy layer. The tabular branch was evaluated on the Wisconsin Diagnostic Breast Cancer dataset (569 cases) with stratified five-fold cross-validation. The calibrated ensemble achieved an accuracy of 98.4% (± 0.9), an F1-score of 0.978, an area under the ROC curve of 0.995 and the lowest Brier score (0.018) of all compared models. Referring the 3.9% of cases with predicted probabilities between 0.3 and 0.7 to clinician review reduced misclassifications from 9 to 4. SHAP analysis identified concave points, perimeter and area of the cell nuclei as dominant, clinically plausible predictors. Image-branch results on histopathology are presented in an illustrative format only, to demonstrate the intended evaluation protocol, and are not measured results. The framework shows how accuracy, calibration, interpretability and privacy can be addressed together in a single clinical decision-support design.