Patch nnU-Net for High-Resolution Semantic Segmentation of Dental X-ray Images

Abstract

High-resolution medical images are becoming more widespread; however, many existing computer vision networks struggle to process these images effectively, often requiring downsampling to lower resolutions. This downscaling leads to the loss of critical details and degrades the segmentation performance. In this study, we propose a method that leverages both high-resolution and low-resolution models by utilizing a patch-based approach to enhance semantic segmentation performance. We aggregated three publicly available dental datasets and applied data augmentation techniques to expand the training set, overcoming the scarcity of dental images. Additionally, we introduced a testing data augmentation method during inference to further improve the segmentation results. Experiments were conducted by training nnU-Net models on un-patched images with different resolutions of 256×256, 350×350 and 1024×512, as well as on patched images of resolution 256×256 extracted from the 1024×512 images using varying stride values. The nnU-Net trained on the high-resolution images of 1024×512 achieved a Dice score of 90.68% and an Intersection over Union (IoU) of 83.44%, outperforming the models trained on lower-resolution images, which obtained a Dice score of 88.93% and an IoU of 80.58% for the 256×256 resolution. However, using the patched nnU-Net model, by decreasing the stride value from 128 to 32 achieved a Dice score of 91.32% and an IoU of 84.12%, surpassing the performance of the model trained on whole high-resolution images.