Construction of an end-to-end regression neural network for the determination of a quantitative index sagittal root inclination.

Abstract

Immediate implant placement in the esthetic area requires comprehensive assessments with nearly 30 quantitative indexes. Most artificial intelligence (AI)-driven measurements of quantitative indexes depend on segmentation or landmark detection, which require extra labeling of images and contain possible intraclass errors. For the initial attempt, the method was tested on sagittal root inclination measurement. This study had developed an accurate and efficient end-to-end model incorporating a convolutional neural network (CNN) based on unlabeled cone-beam computed tomography (CBCT) images for immediate implant placement diagnosis and treatment. The model took pretrained ResNeXt101 as the backbone and was constructed based on 2,920 CBCT images with corresponding angles of the tooth axis and bone axis. The performance of our CNN model was evaluated on a separate test set. Our model exhibited high prediction accuracy in sagittal root inclination measurements, as evidenced by the low mean average error of 2.16°, the high correlation coefficient of 0.915 to manual measurement, and the narrow 95% confidence interval shown by Bland-Altman plots. The intraclass correlation coefficient further confirmed the measurement accuracy of our model was comparable with that of junior clinicians. The model took merely 0.001 seconds for each CBCT image, making it highly efficient. To better understand the model's quality, we visualized our end-to-end CNN model through Guided Backpropagation, Grad-CAM, and Guided Grad-CAM, and confirmed its effectiveness in region recognition. We succeeded in taking the first step in constructing the end-to-end immediate implant placement AI tool through sagittal root inclination measurements without intermediate steps and extra labeling on images.