Uncertainty-aware and interpretable deep learning in cytopathology: toward trustworthy AI-powered diagnostics.

Abstract

Deep learning has shown remarkable success in cytopathology, particularly in cervical cancer screening and thyroid nodule classification. Yet, most models remain "black boxes," offering deterministic predictions without indicating uncertainty or explaining their reasoning, limiting clinical adoption. This review systematically examines uncertainty quantification (UQ) and explainable AI (XAI) techniques for trustworthy deep learning in cytology-the first to do so in this domain. We review state-of-the-art methods such as Bayesian neural networks, Monte Carlo dropout, and deep ensembles for uncertainty estimation, alongside Grad-CAM, attention mechanisms, and surrogate models for interpretability. Applications span cervical, thyroid, urinary, and hematological cytopathology. UQ methods help identify model limitations, flagging cases that need human review, while studies show improved accuracy for high-confidence predictions and better detection of misclassified cases. Visual explanation tools, especially Grad-CAM, highlight diagnostically relevant cellular features that align with pathologists' reasoning. Integrating UQ and XAI fosters effective human-AI collaboration, enabling systems to communicate both confidence levels and diagnostic rationale. This addresses the clinical trust gap, positioning AI as an assistive colleague rather than a black box, ultimately enhancing diagnostic accuracy and efficiency in cytopathology workflows.