Interpretable Multi-Modal Deep Learning for Endodontic Diagnosis and Treatment Planning Using Radiographic Imaging and Clinical Data

Accurate diagnosis and treatment planning in endodontics require the integration of radiographic interpretation and clinical evaluation, making the process highly dependent on clinician expertise and prone to inter-observer variability. Conventional artificial intelligence (AI) approaches in dental imaging primarily rely on single-modality radiographic data, which limits their ability to capture comprehensive clinical context and reduces their effectiveness in real-world applications. To address these limitations, this study proposes an interpretable multi-modal deep learning framework that integrates radiographic imaging, including periapical X-rays and cone-beam computed tomography (CBCT), with structured clinical data such as pain response, pulp vitality, tooth mobility, and patient history for computer-aided diagnosis and treatment-planning support. The proposed framework employs state-of-the-art computer vision architectures, namely EfficientNet and Vision Transformers (ViT), for radiographic feature extraction, while a Multi-Layer Perceptron (MLP) is used to model structured clinical variables. A hybrid attention-based fusion mechanism is designed to integrate heterogeneous imaging and clinical modalities, followed by a classification module for diagnostic prediction and a probabilistic decision-support component for treatment recommendation. To improve clinical transparency and trust, post hoc interpretability methods, including Grad-CAM for radiographic region-based visual attribution and SHAP for feature-level contribution analysis, are incorporated. Experimental results on a clinically curated dataset demonstrate that the proposed multi-modal framework achieves a diagnostic accuracy of 94.8%, outperforming state-of-the-art single-modality models, including ResNet, DenseNet, and EfficientNet, which achieve 87.2%, 90.3%, and 91.5%, respectively. This corresponds to an accuracy improvement of approximately 7.6%, a 3--5% increase in F1-score, and an ROC-AUC of 0.96. In addition, the proposed framework shows a 12.4% improvement in Cohen’s Kappa agreement with expert clinicians, indicating enhanced diagnostic consistency. Overall, the proposed interdisciplinary framework advances dental medical imaging by providing a robust, interpretable, and clinically deployable decision-support system for endodontic diagnosis and treatment planning.