Automatic maxillary sinus segmentation and pathology classification on cone-beam computed tomographic images using deep learning

BackgroundMaxillofacial complex automated segmentation could alternative traditional segmentation methods to increase the effectiveness of virtual workloads. The use of DL systems in the detection of maxillary sinus and pathologies will both facilitate the work of physicians and be a support mechanism before the planned surgeries.ObjectiveThe aim was to use a modified You Only Look Oncev5x (YOLOv5x) architecture with transfer learning capabilities to segment both maxillary sinuses and maxillary sinus diseases on Cone-Beam Computed Tomographic (CBCT) images.MethodsData set consists of 307 anonymised CBCT images of patients (173 women and 134 males) obtained from the radiology archive of the Department of Oral and Maxillofacial Radiology. Bilateral maxillary sinuses CBCT scans were used to identify mucous retention cysts (MRC), mucosal thickenings (MT), total and partial opacifications, and healthy maxillary sinuses without any radiological features.ResultsRecall, precision and F1 score values for total maxillary sinus segmentation were 1, 0.985 and 0.992, respectively; 1, 0.931 and 0.964 for healthy maxillary sinus segmentation; 0.858, 0.923 and 0.889 for MT segmentation; 0.977, 0.877 and 0.924 for MRC segmentation; 1, 0.942 and 0.970 for sinusitis segmentation.ConclusionThis study demonstrates that maxillary sinuses can be segmented, and maxillary sinus diseases can be accurately detected using the AI model.

Journal

BMC ORAL HEALTH

Year

2024

Volume

24

Issue

1

ISSN

1472-6831

DOI

10.1186/s12903-024-04924-0

WoS UID

WOS:001336903600001

Open via DOI