Fuzzy Adaptive Loss Based on Quality and Region Weighting for Medical Image Segmentation

Abstract In medical image segmentation, models are often trained on data influenced by uncertainty, including poor image quality, unclear anatomical structures, and inconsistent annotations. These issues frequently lead to unstable optimization and biased predictions. We addressed these challenges in the context of drug-induced sleep endoscopy (DISE) images, a clinically relevant but difficult domain with high variability in quality and anatomical visibility. In this study, the segmentation target is the airway, whose contour is often uncertain and difficult to delineate accurately. To handle uncertain airway boundaries, we integrated image quality and annotated airway area ratio through a fuzzy rule system. The resulting adaptive weight was then used to adjust the loss during training. This loss function encouraged the segmentation model to focus on diagnostically significant structures while reducing the impact of ambiguous or low-quality images. Performance was evaluated using repeated cross-validation with a 5 $$\\times $$ × 5-fold design and compared against a linear weighting baseline. The fuzzy-guided formulation improved the Dice score from 0.704 to 0.720. In conclusion, clinical attributes such as image quality and airway boundaries are inherently ambiguous. By embedding fuzzy expert knowledge into the training process, we enabled gradient modulation that aligns model learning with semantic uncertainty and improves segmentation performance.