MR–HCP: Morphology-Regularized Hierarchical Conformal Prediction for TEM of Subcellular Ultrastructure

Zahoor Ahmad, Mahmood Alzubaidi, Jens Schneider, Marco Agus, Abeer R. Al-Shammari, Wared Nour-Eldine, Janarthanan Ponraj, John Agcaoili, Samia M. Ltaief

Keywords: Conformal Prediction; Transmission Electron Microscopy; Uncertainty Quantification

TL;DR: MR–HCP is a novel morphology-regularized hierarchical conformal method that delivers calibrated uncertainty (reliable, small prediction sets) for subcellular ultrastructure in TEM images

Abstract: Reliable uncertainty quantification is critical for deploying deep learning models in biomedical imaging, where fine-grained structures often exhibit overlapping morphology and ambiguous boundaries. We introduce Morphology-Regularized Hierarchical Conformal Prediction (MR-HCP), a novel framework that combines hierarchical taxonomies with morphology-aware nonconformity to provide compact prediction sets with rigorous coverage guarantees. Unlike existing conformal methods that operate on flat label spaces and probability-only scores, our approach uses a two-stage super$\to$fine calibration and penalizes morphological deviations from class prototypes. We integrate MR-HCP with a YOLOv11 detector to demonstrate an end-to-end pipeline for single-cell TEM analysis. On a curated neutrophil ultrastructure dataset (around 4.9k annotations), the method achieves near-nominal coverage ($0.937$), small average set size ($1.12$), and high singleton accuracy ($0.932$), significantly outperforming Split CP, Mondrian CP, HCC, and APS baselines. On the public Raabin-WBC dataset, it similarly attains strong coverage-set-size trade-offs compared to conformal baselines. Beyond inference, MR-HCP facilitates semi-automatic annotation and dataset reclassification, systematically refining coarse expert labels into fine categories while transparently quantifying uncertainty. Overall, the framework establishes a morphology-aware hierarchical conformal approach for uncertainty-aware classification in biomedical microscopy and provide a principled basis for extending calibrated set prediction to other hierarchical, morphology-rich biomedical settings.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 5138