Can deep learning improve the interpretation of DPTs? V1
Research type
Research Study
Full title
Can deep learning improve diagnostic accuracy and efficiency when interpreting osseous diseases present on dental panoramic tomograms?
IRAS ID
308048
Contact name
Kawal Rhode
Contact email
Sponsor organisation
King's College London
Duration of Study in the UK
1 years, 6 months, 1 days
Research summary
The dental panoramic tomogram (DPT) is a type of x-ray image used to investigate disease within the jaws. These images are regularly performed and interpreted by dentists, with no input from a specialist radiologist. DPT images are complex and have a large field of view, which results in misdiagnosis and failed identification of significant incidental findings. DPTs are commonly referred for reporting by specialists in Dental and Maxillofacial Radiology (DMFR) but this is not feasible for all DPTs due to the small number of registered DMFR specialists.
Recently, artificial intelligence (AI) has been increasingly implemented into healthcare. Deep learning (DL) is one branch of AI that is particularly well suited to recognising patterns within visual data. DL tools have been developed for diagnostic interpretation of x-ray images and recent studies have shown comparable performance to radiologists when diagnosing pathology within chest x-rays. Furthermore, DL has demonstrated utility in worklist prioritisation of x-ray reporting, for example by automatically flagging computed tomography (CT) images showing suspected intracranial haemorrhage for urgent reporting.
This study aims to assess whether DL can improve the interpretation of DPT images when compared to non-specialist reporters. Additionally, we aim to assess whether DL can assist in the identification of significant pathology, such as cancer, so that these images can be urgently reported by a specialist.
The study will involve retrospective review of DPT images performed as part of direct clinical care at Guy’s Hospital in London. A database of images showing disease will be collected and labelled by DMFR clinicians with a Trust-approved labelling software. A DL tool will then be developed and trained on the database with the aim of identifying, localising and classifying viewable pathology. The DL tool will be tested on a panel of images and diagnostic accuracy will be evaluated against a group of clinicians.
REC name
Yorkshire & The Humber - South Yorkshire Research Ethics Committee
REC reference
22/YH/0134
Date of REC Opinion
9 Jun 2022
REC opinion
Favourable Opinion