Are We Overconfident in Models and Results for Semi-Supervised 3D Medical Image Segmentation?
Researchers have identified a significant overconfidence issue in semi-supervised learning for 3D medical image segmentation. They argue that current methods often confuse prediction confidence with true uncertainty, leading to confirmation bias. Additionally, the lack of dedicated validation sets in many benchmarks encourages the use of test sets for validation, inflating performance estimates and creating an "arms race" of overfitting. To address this, a new framework is proposed that explicitly separates confidence from uncertainty and corrects bias across different data spaces, advocating for more rigorous benchmarking practices. AI
IMPACT Highlights potential overestimation of AI capabilities in medical imaging, urging for more robust evaluation to ensure reliable clinical application.