A few-shot diabetes foot ulcer image classification method based on deep ResNet and transfer learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Diabetology

Background:

• Diabetes foot ulcers (DFU) are a severe complication of diabetes, potentially leading to amputation.
• Current manual DFU classification is time-consuming and lacks accuracy.
• Accurate and early DFU classification is critical for effective patient treatment and management.

Purpose of the Study:

• To develop a few-shot image classification method for Diabetes Foot Ulcers (DFU).
• To leverage deep residual neural networks (ResNet) and transfer learning for DFU classification.
• To address the challenge of limited clinical DFU image data.

Main Methods:

• Employed data augmentation techniques (geometric transformations, random noise) to expand the DFU dataset.
• Conducted comparative experiments selecting various deep ResNet models (ResNet18, ResNet34, ResNet50, ResNet101, ResNet152).
• Utilized transfer learning by fine-tuning pre-trained ResNet models on the augmented DFU dataset.

Main Results:

• Augmented dataset size increased from 146 to 3000 images, improving average DFU classification accuracy from 0.9167 to 0.9867.
• ResNet50 achieved the highest average accuracy (0.9901) and lowest loss (0.1356) among tested models.
• The final DFU classification model, trained with pre-trained ResNet50, reached an average accuracy of 0.9867.

Conclusions:

• The proposed few-shot DFU image classification method demonstrates high accuracy.
• The deep ResNet and transfer learning approach is effective for classifying DFU severity.
• The method is suitable for low-cost, low-computational terminal equipment, aiding clinical DFU diagnosis.