Segmentation of Cemento-Osseous Dysplasias Using an Artificial Intelligence Algorithm

Abstract

Introduction and aims: In recent years, artificial intelligence (AI) has emerged as a powerful tool in medical imaging and in the analysis of complex bone pathologies such as cemento-osseous dysplasias. The aim of this study is to perform segmentation of cemento-osseous lesions using AI algorithms on cone beam computed tomography (CBCT) images and to evaluate the diagnostic performance of a diagnostic AI model designed for the diagnosis of cemento-osseous dysplasias.

Methods: In this study, cone beam computed tomography (CBCT) images taken for various reasons in radiology archive Department of Oral and Maxillofacial Radiology were retrospectively reviewed. As a result of radiographic evaluation, images recorded in the archive with at diagnosis of cemento-osseous dysplasias were determined. Fifty DICOM images were uploaded to the 3D slicer software, and cemento-osseous dysplasias were polygonally labeled and saved in Neuroimaging Informatics Technology Initiative (NIfTI) format. The nnU-Net v2-based automated algorithm for lesion segmentation was developed using the CranioCatch (CranioCatch, Eskişehir) software program using the PyTorch library in the Python framework (v3.6.1; Python Software Foundation). 80% of the data was used for training, 10% for validation and 10% for testing. The results were evaluated according to the criteria of precision, sensitivity, Dice Coefficient, Jaccard Index.

Results: The precision, sensitivity, Dice Coefficient and Jaccard Index for the segmentation of cemento-osseous dysplasias were 0.805, 0.889, 0.839, and 0.730, respectively.

Conclusions: The model we used achieved successful results in cemento-osseous dysplasias segments. The results of this planned study are promising in terms of providing a guidance for physicians in diagnosis.

Clinical relevance: Automated segmentation of cemento-osseous lesions, where radiological images play a critical role in both diagnosis and follow-up, has the potential to enable precise and consistent definition of lesion boundaries and standardize the follow-up process, enabling more reliable data for long-term studies.