Multiclass Bone Segmentation of PET/CT Scans for Automatic SUV Extraction

Abstract

In this thesis I present an automated framework for segmentation of bone structures from dual modality PET/CT scans and further extraction of SUV measurements. The first stage of this framework consists of a variant of the 3D U-Net architecture for segmentation of three bone structures: vertebral body, pelvis, and sternum. The dataset for this model consists of annotated slices from the CT scans retrieved from the study of post-HCST patients and the 18F-FLT radiotracer, which are undersampled volumes due to the low-dose radiation used during the scanning. The mean Dice scores obtained by the proposed model are 0.9162, 0.9163, and 0.8721 for the vertebral body, pelvis, and sternum class respectively. The next step of the proposed framework consists of identifying the individual vertebrae, which is a particularly difficult task due to the low resolution of the CT scans in the axial dimension. To address this issue, I present an iterative algorithm for instance segmentation of vertebral bodies, based on anatomical priors of the spine for detecting the starting point of a vertebra. The spatial information contained in the CT and PET scans is used to translate the resulting masks to the PET image space and extract SUV measurements. I then present a CNN model based on the DenseNet architecture that, for the first time, classifies the spatial distribution of SUV within the marrow cavities of the vertebral bodies as normal engraftment or possible relapse. With an AUC of 0.931 and an accuracy of 92% obtained on real patient data, this method shows good potential as a future automated tool to assist in monitoring the recovery process of HSCT patients.