Domain-adaptive semi-supervised learning for efficient rare pathological lesion detection with minimal annotation

Artificial intelligence for rare pathological lesion detection faces dual challenges: expert annotation scarcity and domain shifts across institutions. Using multi-institutional kidney biopsies from 22 hospitals with 3 scanner types (NDPI, VSI, SVS), we demonstrate that model performance decreases dramatically across domains, with up to 70.3% reduction in detection precision for rare lesions such as crescents and segmental sclerosis (comprising only 2-3% of annotations). We present an approach integrating semi-supervised learning with residual CycleGAN-based domain adaptation, reducing mean Fréchet inception distance between institutions from 55.9 to 20.2 while preserving diagnostic morphology. We identified context-dependent optimal strategies: semi-supervised learning with 50% confidence threshold excelled in same-hospital scenarios (15.2-17.7% improvement for rare lesions), while our combined GAN-Semi-Supervised approach demonstrated superior performance in cross-scanner scenarios between NDPI and VSI formats (up to 63.4% improvement for crescents). This methodology enables robust performance across diverse healthcare settings with minimal expert annotation.