The novel computer-aided diagnosis (CAD) is an approach that is expected to improve the subjectivity of traditional image analysis.  This study aims to assess the threshold and determine the accuracy of computer-aided software for diagnosing proximal surfaces of non-carious posterior teeth.

A simple random sampling of 400 non-carious tooth surfaces from Bitewing radiographs was performed by two experienced observers. The descriptive analysis of threshold value from CAD (Denti.AI) was performed in both digital bite-wing and the available panoramic radiographs of the same tooth surfaces. The reliability of caries diagnostic performance from both radiographic images was calculated by using Cohen’s Kappa Statistics. The inferential statistic was also used for non-normality data and determined by p-value at 0.05.

 CAD showed that 287 (71.8%) tooth surfaces were an absence of proximal carious lesion through bitewing radiographs, while the presence of carious lesions was interpreted in 113 (28.2%) tooth surfaces. On the other hand, CAD showed an almost equal number of negative surfaces (50 or 43.5%) and positive surfaces (65 or 56.5%) through panoramic radiographs. The threshold value (Mean ± SD) of false-positive results of Bitewing and panoramic radiographs were 15.3 ± 12.2 and 34.9 ± 28.0, respectively. The interclass reliability between two types of radiographs, bitewing and panoramic, was -0.055 (p < 0.001), which indicates that there was no agreement in negative result detection in panoramic and bitewing radiographs through computer-aided caries detection software.

 The study showed that computer-aided caries detection software presented a reliable result in ruling out proximal non-carious lesions through bitewing radiographs, which allows the users to considered using the software in proximal caries screening. However, further study should be conducted to determine the comprehension of diagnostic accuracy interpreted by the software in both carious and non-carious lesions.