Medical Image Analysis: SOUSA learns segmentation from sparse annotations

Our team introduced SOUSA in Medical Image Analysis, a unified weakly and semi-supervised framework for medical image segmentation using sparse annotations.

SOUSA is designed for realistic clinical annotation settings, where only a small amount of scribble-level supervision is available and large amounts of unlabeled data remain unused.

Key points:

• Learns jointly from sparse annotations and unlabeled images in a teacher-student framework.
• Combines consistency learning with a multi-angle projection reconstruction loss.
• Outperformed prior weakly and semi-supervised segmentation baselines on multiple datasets.

Paper: Segmentation only uses sparse annotations: Unified weakly and semi-supervised learning in medical images