Convolution Neural Networks and Targeted Fluorescent Nanoparticles to Detect and ICDAS Score Caries

Abstract

Previous work has shown targeted fluorescent starch nanoparticles (TFSNs) can label the subsurface of carious lesions and assist dental professionals in the diagnostic process. In this study, we aimed to evaluate the potential of using artificial intelligence (AI) to detect and score carious lesions using ICDAS in combination with fluorescent imaging following application of TFSNs on teeth with a range of lesion severities, using ICDAS-labeled images as the reference standard. A total of 130 extracted human teeth with ICDAS scores from 0 to 6 were selected by a calibrated cariologist. Then, the same surface was imaged with a stereomicroscope under white light illumination, without visible fluorescence, and blue light illumination with an orange filter following application of the TFSNs. Both sets of images were labeled by another blinded ICDAS-calibrated cariologist to demarcate lesion position and severity. Convolutional neural networks, state-of-the-art models in imaging AI, were trained to determine the presence, location, ICDAS score (severity), and lesion surface porosity (as an indicator of activity) of carious lesions, and tested by 30 k-fold validation for white light, blue light, and the combined image sets. The best models showed high performance for the detection of carious lesions (sensitivity 80.26%, PPV 76.36%), potential for determining the severity via ICDAS scoring (accuracy 72%, SD 5.67%), and the detection of surface porosity as an indicator of the activity of the lesions (accuracy 90%, SD 7.00%). More broadly, the combination of targeted biopolymer nanoparticles with imaging AI is a promising combination of novel technologies that could be applied to many other applications.