33 Computer Vision News Rethinking Semi-Supervised Medical … In this paper, Chenyu proposes and evaluates a novel approach to semisupervised medical image segmentation with limited labels. Recognizing the shortcomings of existing machine learning models, which sometimes lack theoretical grounding in safety-critical scenarios such as medical imaging, this work provides a new theoretical perspective driven by the pursuit of enhancing robustness and efficiency. “Suppose we have a large amount of medical data, and most of the data is unlabeled,” Chenyu poses. “Contrastive self-supervised learning is the predominant approach to train models on a large amount of unlabeled data, but if we directly train models using a contrastive learning framework, they will suffer some model collapse issues.” Preliminary results revealed the model’s tendency to misclassify minority labels, primarily due to class imbalance issues. In medical imaging, different organs account for different numbers of pixels, leading to this imbalance. Ultimately, the contrastive learning framework can go into model collapse, where it misclassifies different pixels from the highfrequency boundaries around the organs. Chenyu’s new sampling method aims to mitigate these issues and improve the reliability of the model. He reports positive outcomes, emphasizing two crucial medical image analysis and machine learning properties: segmentation robustness and label efficiency. “We evaluated our method from these two perspectives,” he explains. “For segmentation robustness, we improved our method compared to the state of the art by 5-10%, and 30% compared to the baseline. Another thing we achieved was an extremely limited label setting.
RkJQdWJsaXNoZXIy NTc3NzU=