Conformal Risk Control for Trustworthy Parasite Egg Detection in Microscopy Stool Images

Abstract

Deep learning object detectors in digital parasitology typically rely on heuristic confidence scores and routinely omit rigorous uncertainty quantification, limiting their trustworthiness for clinical deployment. We introduce a dual-head detector combined with a conformal risk control (CRC) framework to produce effective and risk-controlled predictions with formal, finite-sample risk guarantees. By sequentially conformalizing confidence thresholding, localization, and classification, the pipeline ensures that outputs adhere to a user-specified deployment risk level. Evaluated on a four-class and an 11-class stool parasite egg image dataset, the CRC framework maintains valid risk control, prunes over 92–99% of background proposals, and achieves marginal coverage within finite-sample deviations, with zero spatial expansion under moderate and relaxed requirements. Qualitative analysis confirms that the pipeline adaptively modulates the cardinality of the prediction set in response to instance-level difficulty. Further, an open-set edge case, created by withholding one parasite class during training, shows that CRC suppresses nearly all unknown-class instances, with the rare survivors receiving full prediction sets that explicitly flag uncertainty. By transforming heuristic point estimates into risk-bounded inference, our approach provides a tunable, mathematically rigorous foundation for reliable, deployable AI-assisted microscopic diagnosis of parasitic eggs.