Mamba-Based Identification of Disc Perforation Versus Nonperforation on Temporomandibular Joint Magnetic Resonance Imaging

Objectives To develop and evaluate a hybrid convolutional neural network (CNN)–Mamba deep learning framework for magnetic resonance imaging (MRI)–based expert consensus classification of temporomandibular joint (TMJ) disc perforation vs nonperforation and to assess its preliminary value in a reader study. Materials and Methods We retrospectively assembled 1,695 TMJ MRI scans acquired between June 2018 and December 2023. Reference labels were established as MRI-based expert consensus classifications by experienced TMJ specialists, supported by available clinical documentation. Cases without a reliable consensus classification were excluded. The final dataset comprised 1621 MRI scans. A hybrid CNN-Mamba framework was developed to learn image-based representations from individual 2-dimensional MRI slices. The dataset was split into training/validation and testing sets, and model performance was evaluated at both slice and patient levels. A reader study with 3 experienced dentists compared independent interpretation with model-assisted assessment on the testing set. Results On the validation set, the proposed model showed the highest slice-level performance among the evaluated candidate architectures. At the patient level, sensitivity and positive predictive value were about 0.95 and specificity and negative predictive value were about 0.80, yielding an accuracy near 0.92. With assistance, numerical increases were observed in average reader accuracy (from roughly 0.92 to about 0.97), sensitivity (from approximately 0.91 to about 0.99), and Fleiss’s κ (from 0.760 to 0.893). Heatmap visualisations highlighted the central joint space around the articular disc, providing an anatomically plausible focus for further study. Conclusions A hybrid CNN-Mamba framework showed preliminary potential for MRI-based expert consensus classification of TMJ disc perforation vs nonperforation and may serve as an auxiliary decision support tool in reader assessment. Given the single-centre retrospective design and lack of external validation, prospective multicentre studies are needed before broader clinical application can be considered.