Unsupervised three-dimensional segmentation of scoliotic spines from MR volumes with uncertainty estimation

Abstract

Vertebral bone segmentation from magnetic resonance (MR) images is a challenging task. Due to the inherent nature of the modality to emphasize soft tissues of the body, common thresholding algorithms are ineffective in detecting bones in MR images. On the other hand, it is relatively easier to segment bones from computed tomography (CT) images because of the high contrast between bones and the surrounding regions. This thesis proposes a novel method for a simple thresholding-based segmentation of the scoliotic spine that generates its 3D model by performing a cross-modality synthesis between MR and CT domains. However, this implicitly assumes the availability of paired MR-CT data, which is rare, especially in the case of scoliotic patients. Therefore, the proposed method is completely unsupervised and fully three-dimensional (3D). A 3D CycleGAN model is trained for an unpaired volume-to-volume translation across MR and CT domains. Then, the Otsu thresholding algorithm is applied to the synthesized CT volumes for easy segmentation of the vertebral bones. The resulting segmentation is used to reconstruct a 3D model of the spine. The unsupervised nature of the problem along with the lack of ground truth CT data make it difficult to objectively analyze the performance of the 3D CycleGAN. Therefore, a Bayesian adaptation of CycleGAN is proposed for capturing the uncertainty information in the form of aleatoric and epistemic uncertainties during translation. This enhancement makes the model self-sufficient by providing a measure of the confidence in the model’s predictions while simultaneously making it more interpretable for the Otsu segmentation task. Ablation-study type experiments were run to determine the importance of uncertainty estimation and the improvement in the quality of translated volumes is also shown. The proposed method is quantitatively validated on 46 scoliotic vertebrae in 5 patients with varying degrees of the spinal curvature by computing the point-to-surface mean distance between the landmark points for each vertebra obtained from pre-operative X-rays and the surface of the segmented vertebrae. The study results in a mean distance error of 3.17 ± 1.04 mm. Based on qualitative and quantitative results, it is concluded that the proposed framework is able to obtain good segmentations and 3D models of the scoliotic spines along with the crucial uncertainty information, all after training from unpaired data in an unsupervised manner.

Date	5 May 2021
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Catherine Laporte (Supervisor) & Farida Cheriet (Co-supervisor)