TopoPult-SSL: Gland-Mask-Free Cross-Device Meibomian Gland Segmentation via Self-Distilled Weak Clinical Priors
Researchers have developed TopoPult-SSL, a novel two-stage framework for segmenting meibomian glands across different clinical imaging devices. The first stage adapts existing models using weak clinical priors like eyelid outlines and morphometric ratios, eliminating the need for expensive gland masks during initial training. The second stage refines this with supervised self-distillation when masks are available, achieving competitive results on a benchmark dataset. Notably, the gland-mask-free variant demonstrates significantly higher precision compared to other segmentation models, enabling practical deployment. AI
IMPACT Enables more accessible and precise medical image analysis by reducing reliance on expensive, manually annotated data.