Swin-Unet on Knee X-Ray Images Publisher



Kazemi A ; Zamanirad A ; Esfandiary S ; Najafzadeh E ; Nabian MH ; Farnia P ; Ahmadian A
Source: 32nd National and 10th International Iranian Conference on Biomedical Engineering, ICBME 2025 Published:2025

Abstract

Tibial plateau fractures (TPFs) comprise roughly 1 % of all bone fractures and represent a complex subset of knee injuries with significant clinical implications if not accurately diagnosed and managed. The accurate diagnosis of TPFs from radiographs is challenged by subtle fracture lines and significant inter-observer variability in manual segmentation. To overcome the aforementioned limitations, this study evaluates the performance of a Transformer-based deep learning architecture, Swin-Unet, for automated and precise tibial segmentation. A retrospective dataset comprising 958 anterior-posterior and lateral radiographs from 220 patients with TPFs was curated. Ground truth masks of the tibia bone were manually annotated and validated through a multi-stage review by orthopedic surgeons. Following preprocessing steps, including resizing, adaptive contrast enhancement, and normalization, a 2D Swin-Unet architecture featuring patchbased self-attention mechanisms was trained. Quantitative assessment of the optimized Swin-Unet model on the validation dataset yielded a mean Dice Similarity Coefficient of 0. 8 3 1 4 ± 0.15, a mean Intersection over Union of 0.7374 ± 0.16, and an overall accuracy of 0.9735. Qualitative analysis confirmed the model's ability to accurately delineate tibial boundaries. In conclusion, this study validates the Swin-Unet model as a robust and efficient framework for automated tibial segmentation. By mitigating the challenges of manual delineation, this approach holds significant promise for improving the consistency of orthopedic diagnostic workflows. It serves as a foundation for AI-driven clinical decision support in musculoskeletal imaging. © 2025 IEEE.