Detection and classification of lesions in dental panoramic tomograms

CSC Lead: Mike

Clinical lead: Bethan Thomas

Project Overview Dental panoramic tomograms (DPTs) are common radiographs performed as first-line imaging for a broad range of maxillofacial disease. Within both primary and secondary care, DPTs are interpreted by the requesting dentist, typically without input from a specialist in dental and maxillofacial radiology (DMFR). Specialist reporting is not currently commissioned by the NHS and there are less than 30 registered DMFR specialists nationally, so there is limited scope for second opinion within primary care. Published service evaluations indicate that DPTs referred for specialist opinion contain diagnoses ranging from normal anatomy to malignancy, suggesting that confidence in interpreting DPTs is variable for non-specialists.

DPTs are particularly prone to incidental findings and technical errors, which can obscure interpretation and contribute to misdiagnosis. In some cases, this can lead to missed diagnosis of significant disease, including radiological findings suspicious for malignancy, which leads to delays in appropriate management and subsequent reduced patient outcomes.

A deep learning tool would be clinically useful if it could improve the diagnostic accuracy or efficiency of DPT interpretation by the reporting clinician. Useful output would include differentiation between normal variation and disease. Of the DPTs showing disease, flagging the ROI and suspected clinical behaviour of the lesion (e.g. developmental condition/cyst/benign tumour/malignant tumour) is likely to be more useful than outputting a specific diagnosis. This would assist in workflow prioritisation of time-critical disease e.g. specialist reporting of DPTs with suspected malignancy to streamline appropriate referral from primary to secondary care. The technology would be used by non-specialist clinicians to improve diagnostic accuracy, and could be used by specialist clinicians to improve efficiency.