Original title: Fuzzy Adaptive Loss Based on Quality and Region Weighting for Medical Image Segmentation
Sunny Chung, Pei-Chen Huang, Yun-Ting Lee, Y. Li +3 · September 2026
Computers often struggle to outline human body parts in medical scans. When images are blurry, anatomy is hard to see, or human labels disagree, models make biased predictions and train poorly. Researchers tackled this problem in videos of airways taken during drug-induced sleep endoscopy, where airway borders are very fuzzy and difficult to trace. To solve this, the team built a smart scoring rule. It looks at overall picture quality and the airway's size, then calculates an adaptive weight. During training, this weight adjusts the penalty score, called a loss function. It pushes the model to pay attention to clear, helpful body structures while ignoring confusing or low-quality pictures. This technique raised the model's accuracy score from 0.704 to 0.720, showing that handling ambiguity directly helps computers see difficult medical anatomy much better.
Imagine a teacher grading essays where some essays are written in messy handwriting and others are clear. The teacher normally grades all essays equally, but the messy ones might be unfairly penalized for legibility rather than content. Here, the teacher first checks each essay's handwriting quality and how much of the essay is actually about the topic, then adjusts the grading to focus more on clear, relevant essays and less on messy or off-topic ones. In this paper, the fuzzy rule system evaluates image quality and airway visibility to adjust the training loss, so the model learns more from high-quality images and less from ambiguous ones.
For patients undergoing drug-induced sleep endoscopy, accurate airway segmentation is critical for diagnosing sleep apnea. This work improves that segmentation by reducing the impact of poor-quality images and ambiguous boundaries. Clinicians get more reliable results from automated tools, even when image quality varies.
The method is evaluated only on DISE airway segmentation; its generalizability to other tasks is unknown. The fuzzy rule system requires expert-defined rules for image quality and airway area ratio, which may not transfer to other anatomies or modalities. The improvement in Dice score is modest (0.016), and performance on extremely low-quality images is not separately reported.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Fuzzy Adaptive Loss Based on Quality and Region Weighting for Medical Image Segmentation — Sunny Chung, Pei-Chen Huang, Yun-Ting Lee et al. — International Journal of Fuzzy Systems — doi:10.1007/s40815-026-02334-8
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.