Comparison between observer-based and AI-based reading of CBCT datasets: An interrater-reliability study

To assess the performance of human observers and convolutional neural networks (CNNs) in detecting periodontal lesions in cone beam computed tomography (CBCT), a total of 38 datasets were examined. Three human readers and a CNN-based solution were employed to evaluate the presence of periodontal pathologies in these datasets.

Datasets were acquired with a Veraview X800 L P (JMorita Mfg. Corp., Kyoto, Japan). Three general dentists, previously calibrated by a general principal investigator, read the datasets in 3D MPR mode using Horos(LGPL license at Horosproject.org and sponsored by Nimble Co LLC d/b/a Purview in Annapolis, MD, USA) as a DICOM reader. All pathological changes including vertical bone loss, furcation involvement, and periradicular osteolysis were detected. Furthermore, the same datasets were analyzed automatically by Diagnocat (Diagnocat LLC, Prague, Czech Republic), a deep CNN. Finally, the performance of the dentists and the CNN were compared and evaluated.

The CNN's performance was significantly lower compared to the human readers in the search for different types of lesions. The human observers achieved good to very good interobserver agreement, except for the evaluation of the vertical lesions, which resulted in a moderate agreement.

The CNN used in this study was found to be ineffective in identifying periodontal lesions and was not adequately trained to offer significant assistance in the automated evaluation of periodontal lesions in CBCT datasets.