The Accuracy of Computer Aided Detection of Periapical Radiolucencies on Cone -Beam Computed Tomography Images Using Artificial Intelligence

Overview

A diagnostic accuracy study to assess the accuracy of a newly developed deep learning model in the automatic detection of periapical radiolucent lesions of upper and lower jaws by comparing it with experienced radiologists' opinion, which represents the ground truth. Hypothesis: The null hypothesis is that the results of the deep learning model are as accurate as the radiologists' opinion.

* Study Design: A diagnostic accuracy study * Setting and Location: Retrospective data collection is planned before the index test and reference standard are to be performed. The CBCT data of this study will be obtained from the CBCT data base available at the department of Oral and Maxillofacial Radiology, Faculty of Dentistry, Cairo University, Cairo, Egypt and from available online data set with different CBCT machines. CBCT scans of Egyptian patients who have already been subjected to CBCT examination as part of their dental diagnosis and/or treatment planning will be included according to the proposed eligibility criteria. B) Participants: Based on sample size calculation, a sample of 50 periapical radiolucent lesions of upper and lower different locations in jaw found in CBCT scans. The selection of the scans to be included will be based on the following eligibility criteria. Inclusion criteria: * CBCT scans of maxilla and mandible with good quality free pf periapical radiolucent lesions . * CBCT scans of maxilla and mandible with good quality showing periapical radiolucent lesions. Exclusion criteria: • CBCT images of sub-optimal quality or artifacts / high scatter interfering with proper assessment. C) Variables: • Details about variable CBCT data with periapical lesions were anonymized in DICOM format, and Then, the files will be forwarded to mathematical engineering department faculty of engineering Cairo university for implementing the deep learning model which will include two phases. D) Data Sources / Measurements: • CBCT scans of anonymized retrospective data will be used for research, without the active involvement of patients. Different CBCT machines scans will be used and preliminarily imported into CBCT viewer software program Blue Sky Bio to detect the periapical radiolucent lesions. Localization datasets. 1. Incisors-canines (anterior teeth) Maxillary Mandibular 2. premolars -molars (posterior teeth) Maxillary Mandibular CBCT data with periapical lesions were anonymized in DICOM format, and Then, the files will be forwarded to mathematical engineering department faculty of engineering Cairo university for implementing the deep learning model which will include two phases. Training and validation phase Testing phase Test set separation. Following the completion of the first stage of annotation, a test was separated from the annotated data pool and excluded from all following development activities. Model development dataset. A set for model development purposes formed from the remaining annotated data pool (i.e. not included in the test set) was split into training and validation subsets as it was ft for the task. As the performance of deep learning-based methods heavily relies on a large number of labeled datasets. Existing CNN-based methods (12, 16) first pre-train their model on available online dataset. Therefore, we will use an available online dataset provided by Abdolali et al as data used in their study (18) for research work and then fine tune the network on our collected sample. The exact number of the scans to be used for training can change to avoid underfitting or overfitting in the model, So it will be exactly assigned by the engineering mathematic engineering department faculty of engineering Cairo university. E) Addressing potential sources of bias: No source of bias. The index text will be carried out by a computer program which will be blinded from the ground truth set by 2 well experienced radiologists prior to conduction of this index test. he assessor of the reference standard will not be subjected to the results of the index test as both the gold standard and the index test results will be tabulated and sent to the statistician finally for comparison by a different person rather than the assessors. F) Study Size: A power analysis was designed to have adequate power to apply a two-sided statistical test of the null hypothesis that results of deep learning model are as accurate as the radiologist opinion. By adopting a (95%) confidence interval and by using a specificity value of (88.0%) of the DL group based on the results of a previous study (20) (21) and 100 % for the ground truth: sample size calculated based on specificity was (50) samples. Sample size calculation was performed using Connor equation. G) Sampling strategy: random sampling. H) Quantitative variables I) Statistical methods: sensitivity, specificity, positive predictive value, and negative predictive value will be calculated and graded according to the ranking for diagnostic tests by Leonardi Dutra et al (19) with scores .80% considered excellent outcomes, between 70% and 80% good, between 60% and 69% fair, and ,60% as poor. Teeth without periapical radiolucency served as controls.

Conditions

Eligibility

Outcomes