文献类型：期刊文献

英文题名：Quality-Aware Pseudo-Labeling with a SAM3 Teacher for Label-Efficient Wound Segmentation under Dataset Shift

作者：Huang, Ru[1]; Ren, Tian[1]; Zhou, Zhengbing[3]; Bao, Qichen[2]; Chen, Zijian[4]; Liu, Jiannan[5]; Han, Jing[5]; He, Jianhua[6]; Chu, Xiaoli[7]

机构：[1] School of Information Science and Engineering, East China University of Science and Technology, Shanghai, 200237, China; [2] School of Electronic and Information Engineering, Hangzhou Dianzi University, Hangzhou, 310018, China; [3] Shanghai Sixth People’s Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, 200233, China; [4] Institute of Image Communication and Information Processing, Shanghai Jiao Tong University, Shanghai, 200240, China; [5] Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, 200023, China; [6] School of Computer Science and Electronic Engineering, University of Essex, Colchester, CO4 3SQ, United Kingdom; [7] Department of Electronic and Electrical Engineering, University of Sheffield, Sheffield, S1 3JD, United Kingdom

年份：2026

外文期刊名：SSRN

收录：EI(收录号：20260201152)

语种：英文

外文关键词：Image enhancement - Image segmentation - Medical imaging - Morphology - Photography - Quality control - Statistical tests - Teaching

摘要：Accurate chronic wound segmentation from clinical photographs is important for objective wound assessment, but dense pixel-wise annotation is costly and wound datasets often exhibit substantial shifts in appearance and acquisition conditions. We propose a teacher-student framework for label-efficient wound segmentation, in which a LoRA-adapted SAM3 acts as a controllable pseudo-label generator and Swin-UMamba within nnUNetv2 serves as the student segmenter. To improve pseudo-label reliability, we combine multi-prompt ensembling with a lightweight quality score that reflects prompt stability, morphology plausibility, and teacher-student agreement, and use this score through either hard filtering or soft weighting during student retraining. We evaluate the method under 10\%, 25\%, and 100\% labeled budgets with strict leakage control. Experiments include within-dataset evaluation on wound-photography datasets and auxiliary cross-domain robustness tests on dermoscopy datasets. Beyond the student-only baseline, we compare with Mean Teacher, Cross Pseudo Supervision, and naive pseudo-labeling under the same backbone and budget protocol. Results show that the proposed framework yields consistent gains on wound-photography tasks, while comparisons with naive pseudo-labeling indicate that the improvement is not explained by pseudo labels alone. Hard filtering is generally more robust than soft weighting in more challenging shifted settings. Overall, the proposed framework offers a practical way to improve low-label wound segmentation under dataset shift without changing the deployed student model. ? 2026, The Authors. All rights reserved.