Dual-Attention EfficientNet Hybrid U-Net for Segmentation of Rheumatoid Arthritis Hand X-Rays

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Radiographic imaging segmentation is challenging due to contrast variations, artifacts, and fine anatomical details.
• Existing methods struggle with accurate boundary delineation in medical scans.

Purpose of the Study:

• To develop and evaluate a Hybrid Attention U-Net for enhanced segmentation of radiographic images.
• To improve the recovery of sharp boundaries in medical imaging analysis.

Main Methods:

• Utilized an EfficientNet-B3 encoder paired with a lightweight decoder incorporating CBAM and SCSE attention modules.
• Employed preprocessing techniques including percentile windowing, N4 bias compensation, normalization, and geometric standardization.
• Incorporated sparse geometric augmentations to mitigate domain shift.

Main Results:

• Achieved high performance in hand X-ray segmentation with Dice coefficient of 0.8426 and IoU of 0.78.
• Demonstrated rapid convergence and stable performance during training.
• Ablation studies confirmed EfficientNet-B3's contribution and the benefits of smaller batch sizes (bs=16) for regularization.

Conclusions:

• The Hybrid Attention U-Net offers a computationally moderate, end-to-end trainable solution for radiographic segmentation.
• The model shows potential for extension to multi-class problems and clinical translation.
• It provides a powerful and deployable option for musculoskeletal radiograph segmentation.