Enhancing Burn Diagnosis through SE-ResNet18 and Confidence Filtering

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Computational Pathology

Background:

• Accurate burn severity classification is vital for effective clinical management.
• Current methods struggle to balance precision with real-time performance requirements.
• Deep learning offers potential for automated and accurate burn assessment.

Purpose of the Study:

• To develop a deep learning model for precise and real-time burn severity classification.
• To enhance classification accuracy using an improved ResNet18 architecture with attention mechanisms.
• To optimize the model's performance through adaptive learning rates and advanced optimizer techniques.

Main Methods:

• Implemented a deep learning system with data preprocessing, classification, optimization, and post-processing modules.
• Utilized an enhanced ResNet18 architecture incorporating attention mechanisms.
• Employed adaptive learning rates (cosine annealing, class-specific gradient adaptation) and an improved Adam optimizer.
• Incorporated confidence filtering and weighted aggregation for refined predictions.

Main Results:

• The proposed model achieved a classification accuracy of 99.19% ± 0.12.
• Mean Average Precision (mAP) reached 98.72% ± 0.10.
• Demonstrated high diagnostic reliability and potential for real-time clinical application.

Conclusions:

• The deep learning approach significantly improves burn severity classification accuracy.
• The model effectively balances precision and real-time performance for clinical use.
• This method shows promise for enhancing diagnostic reliability in burn assessment.