Interactive AI assisted pediatric burn assessment based on smartphone images.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Medical Imaging

Background:

• Burn injuries pose a significant pediatric health risk, with current depth assessment relying on subjective visual inspection.
• Objective methods like laser Doppler imaging are limited by cost and portability.
• Scarcity of annotated burn data hinders development of objective assessment tools.

Purpose of the Study:

• To develop an objective, accessible, and accurate method for assessing burn depth in pediatric patients.
• To address the challenge of limited annotated burn data by repurposing pre-trained models.
• To create an interactive tool for clinical diagnosis and dataset annotation.

Main Methods:

• Repurposing pre-trained models (SAM) with minimal fine-tuning for burn assessment.
• Replacing SAM's segmentation head with dense linear regression for continuous depth prediction.
• Utilizing 294 smartphone images from 94 patients, annotated by 9 clinicians, for pixel-level comparison and validation.

Main Results:

• SAM-DR achieved a 0.96 Dice score for wound segmentation, demonstrating state-of-the-art performance.
• Interactive thresholding allowed segmentation of different burn depths, comparable to human expert performance.
• Developed an interactive tool supporting clinical diagnosis and data annotation for burn assessment.

Conclusions:

• SAM-DR offers a non-contact, objective solution for burn assessment using smartphone imagery.
• The method effectively addresses the challenge of scarce annotated burn data.
• The developed tool facilitates both clinical decision-making and efficient dataset creation for burn research.