Filling the Gap in Dental Diagnostics: Tooth-wise Semantic Segmentation of X-Ray Radiographs with UNET-DinoV2
A UNET-inspired architecture with Meta's DINOV2 ViT
Authors: Aimon Benfield-Chand, Aruna Srivastava, Yafqa Khan
Dental and oral diseases significantly impact global health, necessitating advancements in diagnostic technologies. This paper introduces a novel approach to dental abnormality detection using panoramic X-ray radiograph segmentation and mask generation. Leveraging the Tufts Dental Database, we develop a multi-step method that segments dental radiographs into individual tooth crops and subsequently fine-tunes a segmentation model to generate dental abnormality masks. Our approach integrates a UNET-inspired architecture with Meta's DinoV2 pretrained Vision Transformer (ViT) as the encoder, providing a balance between precise abnormality localization and robust feature extraction. We demonstrate the effectiveness of our method by comparing it against a baseline model, highlighting the qualitative and quantitative differences in abnormality detection.
Dental and oral diseases significantly impact global health, affecting around 35% of the population with un- treated tooth decay, 11% with severe gum disease, and 2% experiencing tooth loss [4]. Current diagnostic methods rely heavily on manual examinations, leading to potential inconsistencies and discomfort. This study aims to address these issues by implementing deep learning techniques for more accurate and efficient dental abnormality detection.
- By segmenting the dental radiographs from the Tufts Dental Databaseinto individual teeth crops and subsequently training a segmentation model, we generate dental abnormality masks for individual teeth crops [7].
- This method combines tooth segmentation and abnormality detection, aiming to enhance the precision of abnormality localization while developing a model that learns an abstraction of the features defining a dental abnormality.
U-NET architecture using Meta DINOV2
- The Encoder down-samples the image resolution from (140, 56) to (98, 28)
- The ViT backbone processes the downsampled images into 14 patches each embedded as 384-dimensional vectors
- The Upsampling Network reshapes/up-samples the ViT output back to (98, 28)
- The Decoder further up-samples the image using convolutions, transpose convolutions, and skip-connections from the encoder to produce a segmentation mask
- The best UNET and linear segmentation frameworks achieve comparable test accuracies of 89.9% and 89.5, respectively, and achieve similar results across other metrics.
- While the UNET architecture does not exhibit a significant quantitative advantage over the linear decoder baseline, their qualitative performance differs greatly.
