Suitability of CT and MRI Imaging for Automatic Spine Segmentation Using Deep Learning

This paper examines the suitability of both computer-assisted tomography and magnetic resonance imaging modalities as inputs for automatic human spine segmentation using deep learning algorithms. We conducted the study on two segmentation datasets consisting of scan images and human expert annotated ground-truth segmentation masks of MRI and CT, respectively. In our experiment, we also tested the transferability of the trained algorithms to our in-house dataset containing scans of scoliotic patients in both modalities. We applied two different segmentation algorithms using the U-Net network - standard and patchwise segmentation with rotation averaging for both the CT and MRI dataset.The standard segmentation process yielded more precise and consistent results with a dice coefficient of 0.96 on the CT data and 0.94 on the MRI dataset while achieving a lower inference time of 17ms per one scan. The patchwise approach showed slightly better results when transferred to the in-house dataset containing unseen data during training acquired from different scanning machines. When we consider the smaller size of the MRI dataset, the resulting dice coefficient is comparable across both datasets. Our results show that it is possible to use MRI imaging solely for spine examination and segmentation in cases when we need to visualise also the surrounding tissue and at the same time use automatic segmentation methods for 3D spine model preparation.