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]East China Univ Sci & Technol, Sch Informat Sci & Engn, Shanghai 200237, Peoples R China;[2]Hangzhou Dianzi Univ, Sch Elect & Informat Engn, Hangzhou 310018, Peoples R China;[3]Shanghai Jiao Tong Univ Sch Med, Shanghai Sixth Peoples Hosp, Shanghai 200233, Peoples R China;[4]Shanghai Jiao Tong Univ, Inst Image Commun & Informat Proc, Shanghai 200240, Peoples R China;[5]Shanghai Jiao Tong Univ, Shanghai Peoples Hosp 9, Sch Med, Shanghai 200023, Peoples R China;[6]Univ Essex, Sch Comp Sci & Elect Engn, Colchester CO4 3SQ, England;[7]Univ Sheffield, Dept Elect & Elect Engn, Sheffield, England

年份：2027

卷号：96

外文期刊名：DISPLAYS

收录：;WOS:【SCI-EXPANDED(收录号:WOS:001856756400001)】；

基金：This work was supported by the National Natural Science Foundation of China under Grant 62322114, the Fundamental Research Funds for the Central Universities, China under Grant YG2023LC06, the Natural Science Foundation of Shanghai, China under Grant 20ZR1413800, and the Shanghai Key Laboratory Open Project under Grant STCSM 22DZ2229005. The manuscript has been read and approved by all named authors.

语种：英文

外文关键词：Chronic wounds; Wound segmentation; Quality-aware pseudo-labeling; Teacher-student learning; Promptable segmentation; Dataset shift

摘要：Accurate chronic wound segmentation from clinical photographs is essential for objective wound assessment, but dense pixel-wise annotation remains 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 teacher generates promptable pseudo labels 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 based on prompt stability, morphology plausibility, and teacher-student agreement. The score is used for either hard pseudo-label filtering or soft loss weighting during student retraining. We report the main comparisons under 10%, 25%, and 100% labeled budgets using a fixed fold-0 seed-0 protocol based on a single held-out evaluation fold, together with explicit cross-dataset similarity screening. Experiments cover within-dataset wound-photography evaluation and auxiliary dermoscopy tests. Under the same backbone and budget protocol, we compare the proposed framework with student-only training, Mean Teacher, Cross Pseudo Supervision, and naive pseudo-labeling. The proposed framework improves performance in most wound-photography settings, whereas dermoscopy results are more mixed and reveal remaining challenges in boundary-sensitive shifted cases. These findings indicate that the observed gains are not explained by pseudo labels alone, but depend on reliability-controlled pseudo supervision. Paired case-level bootstrap analysis supports the key wound-photography comparisons under the current single-fold protocol, while multi-fold and multi-seed validation remains necessary to further confirm generalizability. Overall, the framework provides a practical low-label training strategy for the evaluated wound segmentation settings without changing the deployed student model.