A deep learning model for burn depth classification using ultrasound imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Dermatology

Background:

• Accurate burn depth identification is crucial for effective treatment but remains clinically challenging.
• Ultrasound imaging offers a non-invasive method to visualize tissue morphology.
• Deep learning models show promise in analyzing complex image patterns.

Purpose of the Study:

• To develop and validate a deep convolutional neural network (CNN) for classifying burn depth using ultrasound images.
• To assess the model's accuracy in differentiating burn depths, particularly the challenging deep-partial thickness burns.
• To explore the potential clinical utility of AI-assisted burn depth assessment.

Main Methods:

• An encoder-decoder CNN architecture was trained on B-mode ultrasound images of ex vivo porcine skin.
• The encoder was re-trained as a classifier using B-mode images of burned in situ skin.
• Performance was evaluated using 20-fold cross-validation, including accuracy, sensitivity, specificity, ROC, and precision-recall curves.

Main Results:

• The CNN achieved 99% accuracy, 98% sensitivity, and 100% specificity in identifying deep-partial thickness burns.
• High area under the curve (AUC) values of 0.99 (ROC) and 0.95 (precision-recall) demonstrate strong diagnostic performance.
• Post hoc analysis revealed the model utilizes discriminative textural features in ultrasound images.

Conclusions:

• The proposed deep learning model accurately classifies burn depth using ultrasound imaging.
• This AI approach shows significant potential for clinical application in aiding burn assessment.
• The technology could enhance diagnostic capabilities with widely available ultrasound devices.