Evaluating One-Class Representations for Ranking Key Instances at Ultra-Low Witness Rates in Whole-Slide Cytology Images – Swarnadip Chatterjee

Finding key instances in whole-slide cytology images—often containing hundreds of thousands of cells—is a needle-in-a-haystack challenge because the witness rate (WR), i.e., the fraction of malignant instances, is often well below 1%. In this seminar, I will present our work evaluating the effectiveness of one-class representation learning, trained to model normal-cell appearance, for identifying malignant cells under such ultra-low WR conditions. We hypothesised that mixing WRs during training biases models toward high-WR slides, and therefore ran controlled, uniform-WR studies in which prevalence was precisely set (5%, 1%, 0.5%, 0.1%, 0.05%) on a publicly available bone marrow cytomorphology dataset having cell-level annotations. We also evaluated our approach on in-house oral cancer slides—clinically realistic but with WRs that cannot be tightly controlled—to demonstrate external validity and practical constraints. Our findings showed that one-class representation learning reliably surfaces malignant key instances at ultra-low WR, outperforming weakly supervised ( SIL/MIL) baselines that degrade as WR drops. Top-K metrics and WR-sensitivity curves further delineate how performance scales with prevalence. We will conclude this seminar with a discussion of WR-aware evaluation (uniform-WR splits on controllable datasets; prior correction and recalibration for clinical cohorts) and whether explicit handling of prevalence is essential for fair, reproducible comparisons.

About Swarnadip Chatterjee