Chronic Wound Image Augmentation and Assessment Using Semi-Supervised Progressive Multi-Granularity EfficientNet.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Chronic wound assessment relies on standardized tools like the Photographic Wound Assessment Tool (PWAT).
• Deep learning models require large, well-annotated datasets, which are often scarce for specialized medical data like wound images.
• Data augmentation techniques are crucial for improving model performance with limited datasets.

Purpose of the Study:

• To augment a small, imbalanced wound image dataset using semi-supervised learning with a larger unlabeled dataset.
• To develop and evaluate a deep learning model for comprehensive wound assessment based on the augmented dataset.
• To compare the performance of the proposed method against baseline models and prior state-of-the-art approaches.

Main Methods:

• Utilized the Photographic Wound Assessment Tool (PWAT) to define 8 key wound attributes for assessment.
• Employed semi-supervised learning and a Progressive Multi-Granularity (PMG) training mechanism to leverage 9870 unlabeled wound images alongside a 1639 labeled image dataset.
• Applied the EfficientNet Convolutional Neural Network for wound scoring on the augmented dataset.

Main Results:

• The Semi-Supervised PMG EfficientNet (SS-PMG-EfficientNet) achieved average classification accuracies and F1 scores of approximately 90% for all 8 PWAT sub-scores.
• The proposed approach outperformed existing baseline models and showed a 7% improvement over the prior state-of-the-art.
• Generative Adversarial Networks (GANs) for synthetic wound image generation did not enhance wound assessment performance.

Conclusions:

• Semi-supervised learning is a powerful technique for leveraging unlabeled medical data to improve deep learning model performance.
• The SS-PMG-EfficientNet approach demonstrates significant potential for accurate and automated deep learning-based wound grading.
• This study highlights the effectiveness of semi-supervised learning in overcoming data limitations for specialized medical image analysis.