Construction and Evaluation of an AI-based CBCT Resolution Optimization Technique for Extracted Teeth

Introduction

In dental clinical practice, cone-beam computed tomography (CBCT) is commonly used to assist practitioners to recognize the complex morphology of root canal systems; however, because of its resolution limitations, certain small anatomical structures still cannot be accurately recognized on CBCT. The purpose of this study was to perform image super-resolution (SR) processing on CBCT images of extracted human teeth with the help of a deep learning model, and to compare the differences among CBCT, super-resolution computed tomography (SRCT), and micro-computed tomography (Micro-CT) images through three-dimensional reconstruction.

Methods

The deep learning model (Basicvsr++) was selected and modified. The dataset consisted of 171 extracted teeth that met inclusion criteria, with 40 maxillary first molars as the training set and 40 maxillary first molars as well as 91 teeth from other tooth positions as the external test set. The corresponding CBCT, SRCT, and Micro-CT images of each tooth in test sets were reconstructed using Mimics Research 17.0, and the root canal recognition rates in the 3 groups were recorded. The following parameters were measured: volume of hard tissue (V1), volume of pulp chamber and root canal system (V2), length of visible root canals under orifice (VL-X, where X represents the specific root canal), and intersection angle between coronal axis of canal and long axis of tooth (∠X, where X represents the specific root canal). Data were statistically analyzed between CBCT and SRCT images using paired sample t-test and Wilcoxon test analysis, with the measurement from Micro-CT images as the gold standard.

Results

Images from all tested teeth were successfully processed with the SR program. In 4-canal maxillary first molar, identification of MB2 was 72% (18/25) in CBCT group, 92% (23/25) in SRCT group, and 100% (25/25) in Micro-CT group. The difference of hard tissue volume between SRCT and Micro-CT was significantly smaller than that between CBCT and Micro-CT in all tested teeth except 4-canal mandibular first molar (P < .05). Similar results were obtained in volume of pulp chamber and root canal system in all tested teeth (P < .05). As for length of visible root canals under orifice, the difference between SRCT and Micro-CT was significantly smaller than that between CBCT and Micro-CT (P < .05) in most root canals.

Conclusions

The deep learning model developed in this study helps to optimize the root canal morphology of extracted teeth in CBCT. And it may be helpful for the identification of MB2 in the maxillary first molar.