Guiding Neural Networks Toward Better Decisions: A Dental Diagnosis and Treatment Planning Framework and Dataset

Despite the growing availability of anonymized medical imaging datasets, dental data remains largely unavailable due to persistent concerns over person re-identification, posing a barrier to training large-scale models for dental tasks. Unlike commercial solutions, recent academic studies seldom attempt to compose a multi-purpose end-to-end dental framework, primarily due to this data scarcity and the complexity of combining diagnosis and treatment tasks. To address these issues, we present the Ajman University Orthopantomography (AU-OPG) Dataset, a panoramic X-ray dataset annotated for the tasks of teeth detection, diagnosis, and radiographic-evidence-based treatment planning. To the best of our knowledge, AU-OPG is the first publicly described dental dataset that combines diagnostic and treatment-planning labels, and the first to provide tooth-aligned oriented bounding boxes. We also propose a detection-then-classification framework that uses two Masked Autoencoders (MAEs) for tooth and neighborhood feature extraction, respectively. These are combined with a depth-wise neural network that takes three heterogeneous channel inputs, providing positional information that enables the model to implicitly utilize the location of a tooth for its class prediction. The proposed framework achieves competitive performance with test weighted F1-scores of 0.85 and 0.88 for diagnosis and treatment planning, respectively, although the minority treatment categories of filling and root canal remain challenging. This work provides a foundation for open-source radiographic decision-support tools that assist, rather than replace, the dentist’s broader clinical workflow. The dataset and source code are available at https://github.com/akvnn/guiding-neural-nets.