Diagnostic Accuracy of Deep Learning in Medical Image Analysis - A Case Study Using Deep Burns

Abstract

Abstract Purpose Precise evaluation of burn depth is essential for determining the appropriate patient care and surgical requirements. This study aimed to examine a supervised machine learning approach that incorporates dynamic feature selection for differentiating between partial-thickness and full-thickness burns, utilizing deep learning patterns in digital images. Method Four deep learning models (VGG-16, ResNet-50, Xception, and EfficientNetV2L), along with two classifiers (Support Vector Machine and Fully Connected layer), were used to extract features from digital images of burn wounds, implementing dynamic feature selection during the training process. The models were trained using 3-fold cross-validation and tested on an unseen data split. Results The proposed method achieved high prediction accuracy, with the best performance achieved using EfficientNetV2L and SVM, yielding a specificity of 99.38%, sensitivity of 100.00%, precision of 99.35%, and an AUC value of 0.9969. Conclusion The results indicate that the proposed approach, which employs dynamic feature selection, holds potential for clinical effectiveness in objectively assessing burn depths. This technique can aid in informed decision-making regarding patient care and surgical intervention. However, further research is required to investigate its robustness in discriminating various skin wound depths.