SHAL: New AI method cuts annotation cost by 11% for pathology images

Deep learning segmentation of histopathology whole-slide images (WSIs) typically requires massive pixel-level annotations, which are costly and time-consuming. Existing active learning (AL) methods suffer from unreliable uncertainty estimation on partially annotated slides, patch-level acquisition mismatched with slide-level workflows, and unaddressed class imbalance. To solve this, Mahsa Vali and colleagues introduced SHAL (Slide-level Hybrid Active Learning), a patient-level AL framework for multi-class histopathology segmentation. SHAL integrates three components: a foreground-aware strategy to suppress background bias, a stage-adaptive mechanism that hybridizes predictive entropy and epistemic uncertainty across learning stages, and a class-aware strategy prioritizing diagnostically relevant tissue classes.

Evaluated on the TCGA colorectal cancer dataset, SHAL achieved the highest Macro Dice at the full annotation budget (0.846) and reached Dice ≥ 0.80 using only 26% of the budget (50 of 190 slides), while competing methods required 37% (70 slides). Across five external cohorts, SHAL attained the highest mean external Macro Dice (0.815) and the smallest internal-to-external generalization gap (0.025 at Round 3, 0.026 at full budget). This demonstrates that patient-level hybrid uncertainty acquisition reduces annotation cost without sacrificing cross-domain generalization in computational pathology, promising faster and cheaper AI deployment in clinical diagnostics.