A Novel Deep Hybrid Model for Automatic Femoral Stem Classification in Hip Arthroplasty From Radiographs: MSFT-Net With CBAM and Transformer Modules

Accurate identification of femoral stem implants in hip arthroplasty is essential for effective revision surgery, minimizing operative complexity, patient morbidity, intraoperative blood loss, and postoperative recovery time. In cases where prior implant data are unavailable, manual identification is often required, posing significant challenges due to its time-consuming and error-prone nature. To solve this problem, a novel hybrid deep learning architecture that includes a convolutional block attention module and a swin transformer with multi-scale feature fusion from pre-trained architectures DenseNet201, VGG19, and InceptionV3 under the transfer learning paradigm was proposed in this study. The proposed multi-scale feature transformer network was trained and validated on a dataset comprising 1266 anteroposterior (A.P.) hip radiographs of 10 different femoral stem implant types. The proposed hybrid deep learning architecture achieved a training accuracy of 96.7% and validation accuracy of 94.86%, significantly outperforming other baseline models. Compared to state-of-the-art methods, the proposed model achieved an absolute accuracy improvement of 9.5% over VGG19 and 7.4% over DenseNet201 and 8.8% over InceptionV3, demonstrating a significant advancement over existing models in femoral stem classification. The average inference time per image was under 1 second. The experimental results demonstrated that the proposed architecture enhances classification performance while reducing overfitting through attention and transformer-based feature refinement. This automated approach facilitates real-time preoperative implant recognition, thereby streamlining surgical planning, potentially reducing operative costs and duration, and improving clinical outcomes.