Explainable Machine Learning Models for Cancer Classification and Diagnosis: A Comparative Review

Abstract

Cancer diagnosis increasingly relies on machine learning (ML) models trained on cytological, clinical, and imaging data. While these models can match or exceed traditional diagnostic accuracy, their adoption in clinical practice is limited by their “black-box” nature. Explainable artificial intelligence (XAI) methods most notably SHAP, LIME, and Grad-CAM have been proposed to address this gap by attributing predictions to interpretable features or image regions. This review synthesizes five recent studies applying ML, with explainability as a recurring theme, to breast, cervical, and brain-tumor classification and to general cancer-risk prediction, comparing model choice, reported accuracy, and the explainability technique (or its absence) used in each. Across the four cancer-type-specific studies, tree-based ensemble methods (Random Forest, XGBoost, Decision Tree) achieved strong performance (96–97.6%) on tabular cytological and clinical data, while a two-stage convolutional neural network combining LIME, Grad-CAM, and SHAP achieved the highest reported accuracy (99.4%) on MRI-based brain-tumor detection; a fifth, broader ten-algorithm comparison on synthetic risk-factor data reports a consistent pattern (CatBoost, 96.67%) but is explicitly flagged throughout this review as resting on synthetic data and a general-scope publication venue, and is weighted accordingly. The review discusses the trade-offs between model complexity, interpretability, and the model-dependence of explanation methods themselves, the varying evidentiary quality of the underlying literature, and outlines directions for building clinically trustworthy diagnostic tools.