Deep learning framework for bronchoscopic diagnosis of burn inhalation injury

Burn inhalation injury (BII) increases mortality and morbidity in burns patients. Accurate bronchoscopic grading, as the gold standard diagnostic modality, is important for prognostication and to optimize management. However, the most common currently used clinical grading system (abbreviated injury score, AIS) for BII, as a standardized grading method, uses manual judgement of visual features of the tracheobronchial mucosa. This is subjective and has limitations in classification accuracy, and reliability. A better, automated bronchoscopic grading system would have great clinical value. Hence, this study tested the predictive capability of a supervised deep learning technology-based rating method for bronchoscopically diagnosed BII. A pre-trained vision transformer model (ViT) was fine-tuned to automatically grade burn inhalation injury from clinical bronchoscopy recordings of 36 patients (1089 quality-controlled frames) at the London Burns centre. Labelled images were differentiated into training, validation, and test sets (70:20:10). The model was then applied to 2 tasks;1. Identification of the severity grade (modified simple system -none, mild, moderate, severe) and 2. Binary - presence or not of BII. Performance indicators (accuracy, precision, F1 and recall) were measured. Then, the ViT was developed further by transfer learning and data augmentation techniques, and predictive performance retested. Test sets of images in the trained model achieved 98.17 % accuracy, 98.15 % F1 score, 98.29 % precision and 98.17 % recall. For task 2, the enhanced model achieved an accuracy of 98.17 %, F1-score 98.21 %, precision 98.36 %, recall 98.17 %. Compared to traditional human visually graded scoring systems, and even other deep learning model-based studies, our method demonstrated a very promising predictive deep learning framework for application in grading inhalation injuries more accurately.