Development of an explainable artificial intelligence model for Asian vascular wound images

No prior work had resolved the specific challenges of interpreting automated wound assessments within diverse Asian clinical settings. Chronic skin lesions impose substantial economic and health burdens on global populations. Clinicians often struggle with the dynamic and intricate nature of these injuries during routine examinations. While machine learning offers potential for diagnostic support, current systems frequently lack transparency. This opacity hinders the adoption of digital tools in busy medical environments. Explainable models provide a pathway to bridge this gap by clarifying how software reaches specific conclusions. That uncertainty drove researchers to investigate how interpretability might improve trust in automated diagnostic systems. Prior research has shown that integrating such technology requires both high performance and clear decision-making logic.

The researchers aimed to develop an explainable artificial intelligence model specifically for analyzing vascular wound images within an Asian population. This project addressed the significant healthcare burden caused by chronic wounds globally. The team sought to overcome the challenges associated with the dynamic and complex nature of wound assessment. They focused on creating a system that provides transparency to facilitate its acceptance in clinical environments. By developing an interpretable algorithm, the authors intended to bridge the gap between automated analysis and human clinical decision-making. The study specifically targeted the classification of four common wound types and the automatic estimation of physical dimensions. Furthermore, the investigators wanted to provide visual explanations for the reasoning behind each automated output. This effort was motivated by the need for reliable digital tools that clinicians can trust in daily practice.

The research team utilized a registry containing 2,957 images from a Singaporean hospital to train their system. They partitioned this collection into separate training, validation, and testing segments to ensure robust evaluation. Investigators applied oversampling and augmentation strategies to refine the input data before processing. Development involved constructing convolutional and deep learning architectures to handle complex image features. The study measured performance using accuracy metrics, F1 scores, and receiver operating characteristic curves. Experts implemented interpretability techniques to clarify the reasoning behind automated diagnostic outputs. A web-based platform was built to visualize both the wound assessments and the corresponding decision logic. Finally, the authors validated the effectiveness of their approach by testing it on 15,476 unlabeled images.

The model achieved an area under the receiver operating characteristic curve of 0.99 for wound classification with 95.9% mean accuracy. For depth classification, the system reached an area under the receiver operating characteristic curve of 0.97 and 85.0% accuracy. Width and length determination tasks resulted in an area under the receiver operating characteristic curve of 0.92 with 87.1% accuracy. Wound segmentation performance reached an area under the receiver operating characteristic curve of 0.95 and 87.8% accuracy. When tested on 15,476 unlabeled images, the classification confidence score reached 82.8% with 60.6% explainability. Depth classification confidence was 87.6% with 68.0% explainability during these real-world tests. Width and length measurements showed 93.0% accuracy and 76.6% explainability on the unlabeled dataset. Segmentation confidence was 83.9% while the explainability score reached 72.1% for these additional images.

The researchers propose that their algorithm effectively identifies and measures vascular wounds in Asian patient populations. This study demonstrates that high-level accuracy and interpretability are achievable simultaneously in clinical imaging tasks. The authors suggest that their web-based application provides a practical interface for healthcare professionals to review automated findings. These results indicate that transparency features help clinicians understand the logic behind machine-generated assessments. The team claims that their approach could eventually function as a reliable clinical decision support system. Integration into existing electronic health records remains a primary goal for future implementation efforts. The authors conclude that their model maintains consistent performance across both labeled and unlabeled image datasets. This work highlights the potential for explainable technology to enhance the management of complex vascular conditions.