Abstract:

Objective

To establish a global-local feature hierarchical fusion image classification network model，improve the reliability and accuracy of burn and scald wound depth assessment.

Methods

A total of 619 wound images of burn and scald patients who were admitted to the Department of Burn and Plastic Surgery at Ruijin Hospital Affiliated to Shanghai Jiao Tong University School of Medicine were collected.Two burn physicians with more than 3 years of work experience independently annotated the images using the image annotation tool LabelMe， and cross validated with other physicians in the department.The image set was divided into a training set， validation set， and test set in a ratio of 7∶2∶1， and the training set was expanded to obtain 2 598 images.Designed and constructed a global-local feature hierarchical fusion network HFNet for pre-training， learned the basic features of images and transferred them to the collected burn image classification dataset.Improved classification accuracy through parameter optimization and compared the precision， recall， F1 score， and inference time of the HFNet model with ConvNeXt， Swin-Transformer，UniFormer， and BiFormer models to evaluate its performance.

Results

Testing results showed that the precision of the HFNet model in the classification tasks of first-degree， second-degree， and third-degree burn wounds was 93.53%， 94.08%， and 86.52%， respectively， with a mean of 91.63%.The recall rates were 91.99%， 89.89%， and 92.71%， respectively， with a mean of 91.69%.The F1 index was 93.56%， 90.96%，and 90.46%， respectively， with a mean of 91.66%.The average accuracy was 92.75%， 91.94%， and 89.51%，respectively， with an average accuracy of 91.40%.The confusion matrix showed that the accuracy of the HFNet model in the classification tasks of first-degree， second-degree， and third-degree burn wounds was 90%，92%， and 93%， respectively.Compared to models such as BiFormer， the HFNet model achieved higher precision with moderate inference speed， striking a good overall balance between accuracy and computational efficiency.

Conclusion

The HFNet model enhances the accuracy and efficiency of burn depth assessment，providing burn specialists with precise classification information to rapidly determine the severity of burn injuries.Additionally， the model enables the accumulation of high-quality classification data， supporting further model optimization.

Key words: Burns, Deep learning, Adaptive feature fusion, Burn depth recognition