Development and validation of a deep learning model for high-performance segmentation of sacrococcygeal pressure ulcers

Background: Accurate segmentation of pressure ulcers (PUs) is useful for effective treatment planning.Traditional methods often lack the necessary efficiency and precision.While recent studies have explored AI-driven techniques for this task, research on their clinical efficacy remains limited.Methods: This study involved 951 patients with PUs located in the sacral or coccygeal regions.Among them, total of 502 features were extracted from the images to build an AI training dataset.Manual segmentation was conducted on 246 patient photographs using 3D Slicer.The remaining 256 images were segmented using a human-in-the-loop approach, resulting in a fully annotated dataset of 502 samples.A comparative analysis was performed to evaluate the Dice coefficients and execution times between manual and semi-automated segmentation methods.Results: In a comparison between AI-assisted semi-automated segmentation and manual segmentation, all six annotators achieved significantly lower average inference times, with no substantial difference in Dice coefficients (91.31% vs. 89.72%).For the 256 manually segmented photographs, a high average Dice coefficient of 90.9% was observed, with an inference time of 37.42 ms.For the entire dataset, the AI model achieved an impressive average Dice coefficient of 95.1%, with a rapid inference time of 32.45 ms per image. Conclusion:This study highlights the potential of AI for delivering both high accuracy and time efficiency in the segmentation of pressure ulcers.The model shows promise for broad clinical applications, particularly in enhancing the speed and consistency of segmentation in the pressure ulcer lesion.