Сохранить в:
| Главные авторы: | , , , , , |
|---|---|
| Формат: | Recurso digital |
| Язык: | |
| Опубликовано: |
Zenodo
2026
|
| Предметы: | |
| Online-ссылка: | https://doi.org/10.5281/zenodo.20079805 |
| Метки: |
Добавить метку
Нет меток, Требуется 1-ая метка записи!
|
Оглавление:
- PURPOSE: Artificial intelligence is increasingly shaping plastic and reconstructive surgical decision-making. Wound assessment tasks, such as classification, measurement, and prediction, are particularly suited to AI automation. However, the diagnostic accuracy and clinical generalizability of image-based models remain uncertain. This systematic review evaluated the performance of AI models for chronic wound classification. METHODS: A systematic search of PubMed, Scopus, and IEEE Xplore (October 1, 2024) identified 3,537 records. After removing duplicates, 3,112 articles were screened. Eligible studies applied image-based AI to human chronic wounds (diabetic, venous, arterial, or pressure ulcers) and compared outputs with a clinical reference standard. A sub-analysis of 33 models focusing on surgical decision-making for condition classification, extracted for accuracy, AUC, sensitivity, and specificity for infection, ischemia, necrosis, gangrene, and staging tasks. RESULTS: The majority of studies (91%) reported quantitative performance. Accuracies ranged from 50-99.6% for infection and 55-99% for ischemia, with more variable results for necrosis (70-85%). Notably, 27 of 33 studies (82%) utilized the Diabetic Foot Ulcer (DFU2020) dataset or derivatives. Models trained and tested on these datasets achieved the highest accuracies (≥95%), whereas those evaluated on independent hospital images showed more moderate, clinically realistic results (65-86%). Patch-based testing predominated (81%), in which wound images were divided into small tiles to increase sample size. Though a common Computer Vision strategy, it can artificially inflate accuracy between patches from the same wound. Three recent studies introduced transformer-based or hybrid transformer-Convoluted Neural Network (CNN) architectures (DeiT, SimpleViT, and SwinDFU-Net). These models can capture spatial relationships rather than focusing on localized features like conventional CNNs (ResNet or EfficientNet). However, due to limited adoption, their reported accuracies varied substantially (50-96%). CONCLUSION: AI models for wound condition classification demonstrate internal accuracy when evaluated on standardized DFU datasets but exhibit reduced performance on independent clinical images. Heavy reliance on a single dataset family and patch-based validation likely overestimates diagnostic capability by reusing limited image pools and non-independent data. Future research should prioritize enriching the quantity and quality of patient wound datasets to ensure clinical translatability. *Source: https://ps-rc.org/meeting/Program/2026/EP117.cgi*