Swin-EffuseNet: A dual-stream attention-based model combining Swin transformer V2 and EfficientNet-BO for bone...

Area of Science:

• Medical Imaging and Diagnostics
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Bone fractures are common, but manual X-ray interpretation can be error-prone, especially for subtle cases like hairline fractures.
• Accurate and timely diagnosis is crucial for effective fracture treatment and preventing long-term complications.

Purpose of the Study:

• To develop and evaluate Swin-EffuseNet, a dual-stream deep learning framework for robust classification of bone fractures.
• To combine global semantic features with fine-grained local textures for improved fracture detection accuracy.

Main Methods:

• A dataset of 4370 X-ray images was curated from public sources (FracAtlas, Bone Break Classification Dataset).
• Swin Transformer V2 and EfficientNet-B0 were integrated using attention-based fusion to extract semantic and texture features, respectively.
• External validation was performed on Hairline Fracture Detection v2 and Bone Fracture X-ray Simple vs. Comminuted Fractures datasets.

Main Results:

• Swin-EffuseNet achieved 92.8% accuracy, 92.4% precision, 91.6% recall, and 91.9% F1-score.
• The model demonstrated high class-wise accuracies, including 87.9% for Hairline fractures.
• Significant performance improvements were observed compared to individual Swin Transformer V2 and EfficientNet-B0 models, with fast inference (2.8 ms/image).

Conclusions:

• Swin-EffuseNet offers an accurate, efficient, and interpretable solution for intelligent fracture classification.
• The framework shows potential for scalable deployment in diagnostic workflows, improving fracture diagnosis.