Artificial intelligence-enabled staging classification of pressure injuries

Abstract

OBJECTIVE:This study aimed to design an artificial intelligence (AI) tool that can more accurately and objectively identify different stages of pressure injuries (PIs).METHODS:In this study, the authors proposed using AI and computer vision to classify PI images by stage. To this end, the authors implemented a classification network and trained it on a set of PIs images labeled with their stages. This dataset included images from 2 different sources, namely the publicly available Pressure Injury Image Dataset (1091 images), and a private dataset from Ko & ccedil; University Wound Research Laboratory (AY-Lab) (572 images). All images were resized to 224x224 and normalized according to the ImageNet-1K dataset before model input. Various deep learning architectures, including ResNet18, ResNet18-Transformer Encoder Hybrid Model, and DenseNet-121, were used for training and testing. Three-fold cross-validation was used to ensure more robust training and testing. Multiple configurations were tested for each model, and the best-performing configuration was identified. Grad-CAM was applied to visualize attention areas for further evaluation of the model results.RESULTS:After 3-fold cross-validation, ResNet18 outperformed all tested models, achieving an average accuracy of 76.92 +/- 0.92% on the 4-class classification task. The model demonstrated the highest precision of 87.35 +/- 5.54% for Stage 1 and the lowest precision of 64.72 +/- 2.66% for Stage 3.CONCLUSIONS:The results of using the proposed computational approach for PI staging are promising. The AI model can automate PI stage classification, making it a valuable tool for clinic experts.