34 We only require 1% of labels in the medical data to achieve results comparable to the supervised method. These are very good improvements, especially for medical image analysis.” The impact of this work is not confined to medical image segmentation alone; Chenyu reveals that the method is extendable to the broader computer vision community. When tested on three semantic benchmarks, the algorithm consistently outperformed existing methods, achieving state-of-the-art results across the board. The key to these advancements lies in the MONA framework, a contrastive learning framework proposed by Chenyu last year and built on the principles of tailness, diversity, and equivalence. “We tried to address medical segmentation based on these three principles but found that if we randomly or naively sampled the pixels, it resulted in model collapse issues,” he reveals. “We introduced two new sampling strategies to address this. Also, we provided a theoretical guarantee to demonstrate that these sampling strategies can effectively reduce the variance issues and mitigate model collapse.” MONA has already been extended to other computer vision domains, showcasing excellent performance. It initially improved the label setting from 5% to 1% and demonstrated a remarkable 10% performance increase across four datasets in both NeurIPS Accepted Paper
RkJQdWJsaXNoZXIy NTc3NzU=