S-CUDA

S-CDUA: Noisy-Label and Domain-Shift Adaptive Segmentation Algorithm

S-CDUA addresses noisy labels and domain shift in medical image segmentation by learning noise-excluding, domain-invariant representations from supervised source data and adapting to an unsupervised target domain.

Key Features:

• Noisy-Label Cleansing: Detects high-confidence clean samples and identifies corrupted labels for correction and reuse in training.
• Domain Adaptation: Learns domain-invariant features to mitigate distribution differences between source and unsupervised target domains.
• Peer Adversarial Networks: Employs two adversarial networks that exchange high-confidence information to reduce error accumulation and domain gap.
• Supervised Propagation: Enables supervised prediction on target-domain test data affected by domain shift.

Scientific Applications:

• Optic Disc and Optic Cup Segmentation: Validated on REFUGE and Drishti-GS datasets for OD and OC segmentation.
• Spinal Cord Gray Matter Segmentation: Evaluated on a multi-vendor dataset for SCGM segmentation.

Methodology:

The framework integrates noisy-label learning and domain adaptation within deep convolutional neural networks. Two peer adversarial networks are trained to identify reliable clean samples, exchange high-confidence predictions, cleanse noisy labels, and update network parameters using both cleaned and recycled data to achieve robust target-domain generalization.