Automated Wound Image Segmentation: Transfer Learning from Human to Pet via Active Semi-Supervised Learning.

Area of Science:

• Veterinary Medicine
• Medical Imaging
• Computer Vision

Background:

• Wound management is crucial in clinical practice, yet automated solutions for pets are lacking.
• Accurate wound assessment improves diagnosis and treatment effectiveness for chronic pet wounds.
• Current automated wound analysis tools are primarily developed for human use.

Purpose of the Study:

• To develop a novel pipeline for segmenting pet wound images.
• To leverage transfer learning (TL) and active semi-supervised learning (ASSL) for automated dataset labeling.
• To provide guidelines for applying TL+ASSL strategies to image datasets.

Main Methods:

• A pipeline combining transfer learning (TL) and active semi-supervised learning (ASSL) was developed.
• A pre-trained model on human wound images was adapted for pet wound images.
• EfficientNet-b3 U-Net and MobileNet-v2 U-Net models were compared for segmentation performance.

Main Results:

• The proposed TL+ASSL strategy achieved 80% correctly segmented pet wound images after five training rounds.
• The EfficientNet-b3 U-Net model demonstrated significantly superior performance compared to the MobileNet-v2 U-Net model.
• The study confirmed that the number of available samples is critical for effective ASSL training.

Conclusions:

• The developed pipeline offers a viable solution for reducing the time and effort required to generate pet wound image segmentation datasets.
• The TL+ASSL approach enhances the efficiency of creating large, labeled datasets for veterinary applications.
• This work paves the way for improved automated wound assessment tools in companion animal care.