WEM-Mamba: a Wavelet-Enhanced MedMamba framework for rotator cuff tear classification in X-ray images

Rotator cuff tear is a common cause of shoulder pain and functional impairment, and its accurate identification is of great significance for clinical treatment decisions. Magnetic Resonance Imaging (MRI) is generally regarded as the important imaging modality for evaluating soft tissue injuries of the rotator cuff. However, in actual clinical workflows, shoulder X-ray examination is widely used for initial assessment due to its convenience and low cost. Nevertheless, the limited visualization of soft tissues such as tendons on X-ray images makes the identification of rotator cuff tears based solely on X-rays challenging. Existing deep learning methods, such as Convolutional Neural Networks (CNNs), Transformers, and Mamba-based models, primarily focus on spatial domain feature modeling and remain relatively insufficient in utilizing texture and edge information in the frequency domain. Based on the above issues, this study proposes a Wavelet-Enhanced MedMamba Framework (WEM-Mamba) for the task of rotator cuff tear classification in X-ray images. This method introduces the Haar wavelet transform into MedMamba to construct a Wavelet-Enhanced SS-Conv-SSM module (WESS), which extracts multi-band frequency domain features and synergistically fuses them with the global modeling capability of the State Space Model (SSM). Experiments were conducted on a shoulder X-ray image dataset provided by the Sixth Affiliated Hospital of Xinjiang Medical University, and the proposed method was compared with 15 representative models (covering CNNs, Transformers, MLPs, and Mamba variants). The results demonstrate that WEM-Mamba achieves superior classification performance on the current single-center dataset, attaining an accuracy of 0.8950, an F1-score of 0.9309, a precision of 0.9078, a recall of 0.9552, and an AUC of 0.9116, with 14.92 M parameters and 2.04G FLOPs. The above results indicate that, under the data and experimental settings of this study, the introduction of frequency domain enhancement information can provide effective supplementation for rotator cuff tear classification in X-ray images. It should be noted that the conclusions of this study still require further validation on larger-scale, multi-center, and external datasets, and its clinical application value also needs continued assessment in subsequent research.