Mirror, Mirror on the Wall: Automating Dental Smile Analysis in Smart Mirrors with CNN and Diffusion Model

ICML 2024 Workshop ML4LMS

Keywords: Machine Learning, Internet-of-Things, Diffusion Model, dental smile analysis

TL;DR: Automating dental smile analysis using a CNN trained on real and generated data with a vision to be integrated into an innovative Internet-of-Mirrors network of smart mirrors.

Abstract: This paper presents a smart diagnostic framework for dental smile analysis. To accurately and efficiently identify esthetic issues from a single image of a smile, a convolutional neural network (CNN) was trained. To overcome the limitations of scarce data, a diffusion model was employed to generate dental smile images in addition to manually curated data. The CNN was trained and evaluated on three datasets: all real images, all generated images, and a hybrid dataset of equal proportions of real-to-generated images. All three models demonstrate accuracy significantly above the baseline in detecting excessive gingival display, unlocking a novel diagnostic method in smile analysis. Notably, the hybrid model achieved the highest accuracy of 81.61\% ($p$ value $<$ 0.01), highlighting the effectiveness of generative data augmentation for machine learning. The proposed solution could be part of a standalone home-deployed smart mirror or connected to a network of an innovative Internet-of-Mirrors to facilitate patient-dentist communication.

Submission Number: 27