Adversarial co-training for semantic segmentation over medical images

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Semi-supervised learning for medical image segmentation relies on labeled and unlabeled data.
• Labeling medical data is costly and time-consuming.
• Distribution shift in unlabeled data (e.g., from different sources) can negatively impact model training and lead to confirmation bias.

Purpose of the Study:

• To investigate the impact of distribution shift on semi-supervised semantic segmentation of medical images.
• To develop robust methods to improve model performance in the presence of distribution shift.
• To enhance the practical utility of semi-supervised learning for medical image segmentation.

Main Methods:

• Introduced adversarial training into a co-training process to simulate and mitigate distribution shift.
• Used adversarial perturbations to enhance model robustness against distribution shifts during supervised training.
• Employed co-training with two sub-models trained on different data views to integrate diverse knowledge and avoid confirmation bias.

Main Results:

• Achieved a Dice Similarity Coefficient (DSC) of 87.37% on the ACDC dataset using only 20% of labels, comparable to using 100% labels.
• On the SCGM dataset, achieved a DSC of 78.65% with 6.5% of labels, outperforming prior state-of-the-art by 10.30%.
• Demonstrated superior robustness against distribution shifts in medical imaging scenarios through extensive experiments.

Conclusions:

• The proposed adversarial co-training approach effectively handles distribution shift in medical image segmentation.
• The methods significantly improve model robustness and performance, enabling high accuracy with minimal labeled data.
• Empirical results validate the effectiveness for practical applications in medical imaging.