Validation of Conformal Prediction in Cervical Atypia Classification

TL;DR

This study validates conformal prediction for cervical atypia classification using multi-expert annotations, revealing that traditional coverage metrics overstate practical performance when assessed against expert labels. It compares three CP approaches (LAC, APS, RAPS) across three CNN backbones, showing high true-class coverage but frequent inclusion of extraneous classes and poor exact alignment with expert sets. The work demonstrates that CP methods better capture aleatoric uncertainty (data ambiguity) than epistemic uncertainty (OOD), with model-dependent behavior in OOD scenarios. The findings underscore the need for cautious interpretation of CP outputs in clinical contexts and call for improved evaluation frameworks that align prediction sets with human expertise and real-world decision-making.

Abstract

Deep learning based cervical cancer classification can potentially increase access to screening in low-resource regions. However, deep learning models are often overconfident and do not reliably reflect diagnostic uncertainty. Moreover, they are typically optimized to generate maximum-likelihood predictions, which fail to convey uncertainty or ambiguity in their results. Such challenges can be addressed using conformal prediction, a model-agnostic framework for generating prediction sets that contain likely classes for trained deep-learning models. The size of these prediction sets indicates model uncertainty, contracting as model confidence increases. However, existing conformal prediction evaluation primarily focuses on whether the prediction set includes or covers the true class, often overlooking the presence of extraneous classes. We argue that prediction sets should be truthful and valuable to end users, ensuring that the listed likely classes align with human expectations rather than being overly relaxed and including false positives or unlikely classes. In this study, we comprehensively validate conformal prediction sets using expert annotation sets collected from multiple annotators. We evaluate three conformal prediction approaches applied to three deep-learning models trained for cervical atypia classification. Our expert annotation-based analysis reveals that conventional coverage-based evaluations overestimate performance and that current conformal prediction methods often produce prediction sets that are not well aligned with human labels. Additionally, we explore the capabilities of the conformal prediction methods in identifying ambiguous and out-of-distribution data.

Figures (8)

• Figure 1: Sample tiles, with their corresponding annotations from four expert annotators and conformal prediction sets generated for a trained model. The output prediction sets perfectly cover one or more of the experts' annotations. However, they also usually add extraneous labels and fail to mirror the disagreement between annotators, as seen in the first three examples. The last prediction set correctly outputs the experts' annotations.
• Figure 2: User interface of the web-based annotation platform for labelling AI-generated regions of interest (ROIs) on whole slide images. A specific ROI is highlighted, and a predefined list of options offers the diagnostic categories, such as NILM, LSIL, HSIL, and Artefact.
• Figure 3: A sample input tile (shown on the left) and its Out-of-Distribution variants generated by adding a Gaussian noise at different $\sigma$ values.
• Figure 4: A summary of classification coverages and mean F1 scores of all conformal prediction methods for $\alpha \in \{0.05, 0.1, 0.15$, $0.2\}$.
• Figure 5: Highlight of individual ground truth categories accurately identified by conformal prediction sets ($\alpha = 0.05$). For each of the tiles, we expected the perpendicular red lines to overlap with the light blue lines across all the categories, which ended up not being the case.