For years, assessing the extent of burns has relied heavily on the physician’s visual judgment. However, due to the variability in human observation, different doctors often produce widely varying assessments of the same burn image, leading to inaccuracies in treatment dosage calculations, which can significantly affect patient outcomes. Factors such as wound color and surrounding tissue swelling often skew the visual assessment. Furthermore, irregularly shaped burns or those in concealed areas of the body make it even harder to evaluate accurately using traditional methods.
To address this issue, Dr. Che-Wei Chang and his research team at Far Eastern Memorial Hospital have developed an AI-powered burn diagnosis platform that combines computer vision with deep learning. This system, trained using tens of thousands of burn images, allows AI to accurately identify burn areas. Unlike traditional visual methods, the AI system is objective, precise, and consistent, performing pixel-level segmentation of burn areas without being influenced by subjective factors. It calculates the percentage of the burn area based on the segmentation results, offering more precise burn assessments.
Deep Learning Unveils Burn Diagnosis Capabilities
The research team’s approach can be broken down into three key phases:
Initial Phase: The AI platform was built on a U-Net architecture using ResNet101 as its backbone. The team gradually integrated Mask-RCNN’s region of interest (ROI) network and DeeplabV3+’s atrous convolution to reduce information loss during computations, thereby improving the model's performance.
Optimizing Annotation Processes: To minimize discrepancies in burn image annotations by experts, the team used superpixel segmentation technology, dividing images into texture-similar blocks, allowing experts to select depth instead of manually annotating burn areas. Once over 6,000 images were collected, the model began to self-annotate the areas, with experts making fine adjustments, significantly reducing the time needed for annotation.
Model Lightweighting: To ensure the large AI model could run on mobile and other edge devices, the team restructured the model using depth-wise convolution, which increased computation speed while maintaining a high level of accuracy.
AI Burn Diagnosis: A Promising Future
The AI system has already been successfully implemented at Far Eastern Memorial Hospital, with clinical trials showing that its diagnosis results are more accurate and consistent compared to traditional physician evaluations.
Emergency Room: In emergency settings where patients are in critical condition, rapid and accurate burn area assessment is essential for developing optimal treatment plans. The AI system provides fast, reliable results to support physicians’ decision-making.
Burn Care Units: In burn care wards, the AI system helps monitor wound healing and allows physicians to adjust treatment plans in real-time, improving therapeutic outcomes.
Community Hospitals: In resource-limited community hospitals, the AI system can serve as a diagnostic support tool, enhancing diagnostic accuracy and providing critical support to local physicians.
Future Outlook
Dr. Che-Wei Chang emphasized that the development of the AI burn diagnosis platform is a successful example of integrating medical imaging analysis with artificial intelligence. This breakthrough offers new hope for burn treatment and serves as a model for other areas of medical development. The research team plans to further explore the integration of the AI system with other medical imaging modalities, such as infrared thermal imaging and 3D scanning, to provide more comprehensive burn information. Additionally, they aim to develop telemedicine and personalized treatments, providing timely diagnosis and care to patients in remote areas. As AI technology continues to advance, more AI-based medical applications are expected to emerge, contributing significantly to human health and well-being.
Resource (mandarin): 補強目測法不足 AI判別燒燙傷搶救更快更精準
