Pretrained Neural Networks for Segmentation of X-Ray Microcomputed Tomography Images of Bones and Teeth

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Volumetric imaging using X-ray microcomputed tomography (µCT) is a mainstream method for 3D structural studies of bones and teeth. Quantitative analysis of mineralized features in general relies on accurate segmentation. The larger the image and the higher the resolution, the greater the incentive to use (semi)automated high-throughput segmentation approaches, primarily using deep learning-aided methods. In this study we pretrained and evaluated two U-Net segmentation models, called Universal Bone and Universal Jaw, using a diverse 3D image library from contributing colleagues. Our assumption was that pretraining a U-Net on skeletal and dental features would result in higher accuracy and throughput, in comparison with a naïve U-Net trained by an end-user on a limited local dataset. The Universal Bone binary segmentor pretrained on a relatively heterogeneous subset of images showed satisfactory generalization capacity, and was applicable to unseen images with minimal tuning. The Universal Jaw four-class segmentor (background, bone, dentin, enamel) was pretrained on a narrower range of images and required rather extensive retraining on manually retouched image segmentations. The performance of pretrained segmentation U-Nets is defined not only by the number of segmentation classes (the fewer the better), similarity of the unseen image spacing to the training set (the closer the better), but even more so by class balance. For particularly large 3D images, extensive manual retouching and re-training was nevertheless valuable attributable to the high throughput of the pretrained segmentor. Despite achieving partial success in pretraining specialized segmentation models, we advocate for future concerted efforts to merge and expand image collections, share pretrained models and utilize expert segmentations for benchmarking.