AI-Driven Orthopedic Implant Identification in Indian Clinical Practice: A Dynamic Cross-Attention Swin Transformer...

This paper proposes a novel Dynamic Cross-Attention enabled Cross-Swin Transformer approach for identifying orthopedic implants efficiently.

Initially, low-dimensional features are captured by employing Hybrid Patch Embedding mechanism, while the Cross-Swin transformer constructs a hierarchical feature representation. Then, Linear Multi-head Self-Attention minimizes computational complexity and broadens the receptive field for identifying large-scale features. After that, the Efficient Channel Attention strategy facilitates cross-channel communication and captures inter-channel dependencies effectively, thus avoiding dimensionality reduction. Additionally, the individualized trade-off among local convolution and global attention is maintained by Adaptive mixture units. The multi-dimensional features are effectively fused by Attention Feature Fusion Unit for optimizing network efficiency. Furthermore, an improved genetic algorithm optimizes hyper-parameters, employing chaotic opposition and Tabu search algorithms to balance global and local optimization.

Overall, the proposed approach identifies orthopedic implants with simple calculations and attains greater accuracy of 99.03% compared to prior orthopedic implant identification systems in terms of some common assessing measures.

The research findings show how well the proposed technique works to identify the manufacturer and the model of orthopedic implants accurately, aiding orthopedic surgeons in the pre-operative planning of revision surgery.

The online version contains supplementary material available at 10.1007/s43465-025-01432-3.