AutoMated Burn Diagnostic System for Healthcare (AMBUSH)

BACKGROUND: In the United States (US), about 1.25 million people are treated each year for burns, and 40,000 are hospitalized for the treatment of these injuries resulting in high medical costs, approximately $7.9 billion per year. Early assessment of burn depth considered a predictor of pathological scarring that occurs in 30%-91% of burn injuries, and prioritizing burns that require surgical intervention is a critical task. However, it continues to be an open clinical challenge. We sought to develop a high accuracy automated system, that relies on multimodal Harmonic B-mode ultrasound (HUSD B-mode) and Tissue Elastography imaging (TEI), to classify burn pathology using novel techniques based on machine learning and artificial intelligence (AI). METHODS: Burn wounds of different degrees (superficial, partial, and full thickness; n=2 each per pig; size 2“x2”) were created on the dorsum of female domestic pigs (70-80lbs) (n=6 pigs) using a standardized burner. Burn wounds were treated with the same dressing. Progression of burn wounds was followed by non-invasive imaging using digital photographs, HUSD B-mode, and TEI videos at day 0 - postburn, and on days 3, 7, 14, 21, 28, 35 and 42 postburn. Burn depth was validated by histopathological analysis and results were compared with US-acquired data at different time points. State-of-the-art deep learning methods to analyze images and videos such as convolutional neural networks (CNNs) were employed. These features were used to train task-specific networks. In the case of depth classification, the classifier was further enhanced using traditional computer vision features. RESULTS: Burns of different degrees were successfully created on all the pigs. HUSD B- mode and TEI showed characteristic biomechanical and biological response patterns unique to the different degrees of burn which was validated by H&E staining. Histological pattern graded the burn injury from superficial involving only epidermal layer to the full thickness burn involving all skin and subcutaneous layers. Data labelling, segmentation and augmentation was done and fed into the AI system. Our system was able to classify burn wounds with a mean accuracy greater than 90%. In burn segmentation, our system achieved a mean global accuracy greater than 0.87. Further, we calculated a mean intersection over union (IoU) score of ~0.8. These scores represent a statistically significant improvement over our baseline segmentation model. Critically, this part of our system presented a clear and human-readable masks to understand the surface of burn wounds, allowing a high degree of explainability often required to interpret AI-produced results. CONCLUSIONS: This work presented elements of an autonomous AI system to analyze and predict burn depth via texture-based image processing algorithms using multiple common medical modalities.