Advancing Interpretable AI for Clinical Decision Support and Image-Driven Cancer Detection: A Review of Reliability, Transparency, and Diagnostic Performance

Abstract

Explainable AI has been studied separately for cancer classification and for other image-based clinical decision support tasks such as pneumonia detection, but a unified account of what these two literatures share, and where they diverge, is largely missing. This review synthesizes seven studies to build that account. Two systematic reviews anchor the synthesis: one covering 46 explainable-AI cancer-imaging studies finds that Grad-CAM dominates as an explanation method (used in 34 of 46 studies) while only 10.9% of studies evaluate their explanations quantitatively rather than through subjective visual assessment; a second, broader review of 133 studies spanning X-ray, CT, MRI, PET, ultrasound, and pathology across pneumonia, cancer, and other conditions identifies standardization and clinical-workflow integration as the field's most persistent unresolved barriers, regardless of which disease is being diagnosed. Applied studies on both sides of this review's title support that convergence: a hybrid CNN explained via Grad-CAM++ reaches 97.14% accuracy classifying breast-cancer ultrasound images, a Random Forest classifier reaches 96% accuracy on a cytological cancer benchmark, and, on the clinical-decision-support side, a ResNet50 model compared across four interpretability methods reaches 91% accuracy detecting pneumonia from chest X-rays with the best-performing explanation method (layer-wise relevance propagation) adding no accuracy cost. Two further studies from the same research cluster propose, respectively, a multimodal cancer-diagnosis framework and a MobileNet-based pneumonia detector; both are discussed with an explicit, disclosed limitation, since their full quantitative claims could not be independently verified beyond their confirmed titles and a general-scope publication venue this review examined directly. Read together, these seven studies suggest that a genuinely unified explainable-AI framework for clinical decision support is less a matter of building one model that does everything than of applying the same standard of quantitative, clinically validated explanation across every disease and modality a healthcare system needs to diagnose a standard this literature has not yet consistently met even within cancer imaging alone.