Investigation reveals novel frameworks for enhancing image segmentation in medical AI, addressing data scarcity and architectural challenges.
Key Points
The aim is to develop advanced models for medical image segmentation that effectively leverage both labeled and unlabeled data while addressing inherent architectural and data scarcity challenges.
Introduced CS-Unet, a Transformer network that incorporates convolutional blocks.
Developed a Multi-Scale Cross Supervised Contrastive Learning framework for collaborative training of CNNs and Transformers.
Proposed a certainty-guided contrastive learning strategy to reduce the impact of noisy pseudo-labels.
Implemented the Contrastive Cross-Teaching with Registration framework to integrate spatial registration into the learning process.
CS-Unet outperforms existing models on multi-organ and cardiac datasets without requiring pre-training.
MCSC demonstrates improved segmentations by training CNNs and Transformers together using multi-scale contrastive losses.
The certainty-guided approach shows robustness against noisy label impacts, improving reliability of segmentation.
CCT-R enhances learning efficiency and accuracy using anatomical priors and achieves state-of-the-art performance in segmentation benchmarks.