Enhanced skin burn assessment through transfer learning: a novel framework for human tissue analysis

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Dermatology

Background:

• Visual inspection is standard for burn evaluation, but may miss subtle burn severity.
• Inaccurate burn severity assessment can lead to delayed treatment, scarring, organ failure, and death.
• Traditional clinical methods struggle with early, effective prediction of burn wound severity.

Purpose of the Study:

• To address the gap in artificial intelligence research regarding transfer learning for burn severity assessment.
• To improve the performance of machine learning models in classifying burn degrees using transfer learning.
• To develop a computer-aided diagnosis tool for prompt and precise burn damage evaluation.

Main Methods:

• Utilized transfer learning techniques with deep learning (DL) models, including a basic convolutional neural network (CNN) and seven distinct transfer learning models.
• Classified burn images into first, second, and third-degree burns, as well as healthy skin, using a fully connected feed-forward neural network.
• Trained and tested models on combined datasets from Kaggle 2022 and the Maharashtra Institute of Technology open-school medical repository.

Main Results:

• The basic CNN model achieved an accuracy of 93.87%.
• The VGG-16 transfer learning model demonstrated the highest accuracy at 97.43%.
• The DenseNet121 model followed closely with an accuracy of 96.66%.

Conclusions:

• The proposed CNN-based approach incorporating transfer learning significantly enhances the accuracy of burn severity assessment compared to basic CNN models.
• This AI-driven method can assist healthcare professionals in timely and accurate burn damage evaluation, leading to improved treatment outcomes.
• The study highlights the effectiveness of transfer learning in improving machine learning model performance for medical image analysis in burn care.