AI for Cytopathology

Abstract:

The intersection of Deep Learning (DL) and pathology has gained significant attention, encompassing cell classification, detection, segmentation, and whole-slide image (WSI) analysis. Further works at this intersection have increasingly focused on integrating uncertainty quantification (UQ) with DL methods for pathology to address their occasional unreliability in clinical settings. Conformal Prediction (CP) is one of the UQ methods deployed for various medical settings including pathology. CP methods are computationally efficient and offer user-defined coverage guarantees to generate prediction sets that include the true label. However, CP methods lack inherent control over the compositionality of prediction sets, which restricts their clinical utility. This study presents a novel hinge loss-based training method for the underlying models used in CP methods. This approach aims to provide effective control over the compositionality of prediction sets, aligning more closely with the specific needs of pathologists. We evaluate the effectiveness of this training approach using three application-specific metrics tailored to enhance the integration of CP methods into clinical pathology workflows. Our results show that the Hinge Loss-based training approach outperforms the traditional Cross-Entropy method across all evaluation metrics, effectively managing the compositionality of conformal prediction sets.