Abstract: Current clinical diagnosis of temporomandibular joint disorder (TMD) mainly relies on the manual interpretation of magnetic resonance imaging (MRI) results by dentists, which suffers from high subjectivity and low efficiency. Although deep learning technologies have been applied to the automatic diagnosis of TMD, most existing methods are limited to single-modal imaging data. They ignore the critical diagnostic prior knowledge contained in non-imaging information (i.e., clinical tabular data) and fail to effectively model the long-range global dependencies of MRI images. To address these issues, this paper proposes a Mamba-based multi-modal automatic diagnosis method for temporomandibular joint diseases, which integrates MRI images and clinical tabular data to improve the accuracy and generalization ability of diagnosis. Firstly, a Mamba-based image encoder adopts cascaded dual-branch state space blocks (DSSB) to simultaneously capture local details and long-range global context of MRI images, so as to fully extract discriminative image features. Meanwhile, a BERT-based text encoder processes clinical tabular data to acquire personalized diagnostic features of patients. Subsequently, a text-guided attention fusion (TAF) module leverages the optimal transport attention mechanism to fuse the above complementary features and realize effective multi-modal feature interaction. Finally, the fused multi-modal features are input into the prediction layer to achieve automatic classification of three types of lesions. Experimental results on a real clinical TMD dataset show that the proposed method achieves an area under the curve (AUC) of 0.918, an accuracy of 0.780, and an F1-score of 0.774, which are significantly superior to those of existing single-modal and multi-modal diagnosis methods.