Usability of a deep learning platform for detecting radiographic bone loss and furcation involvement

Assessing radiographic bone condition is important for periodontal diagnosis. The accuracy of radiographic interpretation depends highly on a clinician's experience and knowledge. This study aimed to develop a deep learning-based online platform that aids clinicians in diagnosing periodontitis based on periapical radiographs and to evaluate the platform's usability using a validated survey.

An end-to-end deep learning-based platform was developed to measure radiographic bone loss (RBL), identify furcation involvement, and suggest a case-level periodontal diagnosis from periapical radiographs. The platform's performance was evaluated based on 100 clinical cases. Dental students, residents, and dentists evaluated the platform's usability by completing a survey with the System Usability Scale (SUS) and clinical application questions. Comparisons of survey scores between groups were analyzed by one-way analysis of variance (ANOVA).

In total, 6,552 intraoral radiographic images were used to develop the deep learning models integrated with the online platform. These images had 662 sites with vertical bone loss and 864 radiographic furcation involvement defects present. The platform's accuracy for Extent (0.86), Stage (0.87), and Grade (0.97) of the clinical cases' diagnosis was generally high. The mean scores of the SUS questions for dental students (n = 20), residents (n = 20), and dentists (n = 20) were 79.13 ± 12.01, 74.13 ± 16.17, and 80.75 ± 15.24, respectively (p = 0.333).

This reliable deep learning-based platform can be potentially utilized as an adjunct diagnostic tool and is highly accepted by care providers.

Assessing bone conditions around teeth through radiographs is essential for accurate periodontal diagnosis. The effectiveness of radiographic interpretation relies heavily on the clinician's experience and expertise. It will be beneficial to have an artificial intelligence-driven platform that can provide standardized and reliable radiographic bone assessment results. This study used 6,552 dental radiographs to develop multiple deep learning models. A platform integrated with these deep learning models was created to assess bone conditions around teeth and suggest a case-level periodontal diagnosis based on dental radiographs. A validated survey with a score ranging from 0 to 100 was used to collect care providers' feedback after testing the platform. Our results showed that the platform demonstrated high accuracy in all diagnostic categories, including extent, stage, and grade, of clinical cases. The mean survey scores for 20 dental students, 20 dental residents, and 20 dentists were 79.13 ± 12.01, 74.13 ± 16.17, and 80.75 ± 15.24, respectively. Overall, the deep learning-based diagnostic platform can reliably assess radiographic bone conditions and generate reliable case-level diagnoses. The usability of this platform is well accepted by healthcare providers. As an adjunct diagnostic tool, this platform can assist clinicians in making accurate periodontal diagnoses, leading to more effective treatment planning.