Automatic Bone Structure Segmentation of Under-Sampled CT/FLT-PET Volumes for HSCT Patients
| dc.contributor.advisor | Havlicek, Joseph | |
| dc.contributor.author | Carson, Brandon | |
| dc.contributor.committeeMember | Cheng, Samuel | |
| dc.contributor.committeeMember | Zheng, Bin | |
| dc.date.accessioned | 2021-08-03T21:24:45Z | |
| dc.date.available | 2021-08-03T21:24:45Z | |
| dc.date.issued | 2021-08-05 | |
| dc.date.manuscript | 2021-07 | |
| dc.description.abstract | In this thesis I present a pipeline for the instance segmentation of vertebral bodies from joint CT/FLT-PET image volumes that have been purposefully under-sampled along the axial direction to limit radiation exposure to vulnerable HSCT patients. The under-sampled image data makes the segmentation of individual vertebral bodies a challenging task, as the boundaries between the vertebrae in the thoracic and cervical spine regions are not well resolved in the CT modality, escaping detection by both humans and algorithms. I train a multi-view, multi-class U-Net to perform semantic segmentation of the vertebral body, sternum, and pelvis object classes. These bone structures contain marrow cavities that, when viewed in the FLT-PET modality, allow us to investigate hematopoietic cellular proliferation in HSCT patients non-invasively. The proposed convnet model achieves a Dice score of 0.9245 for the vertebral body object class and shows qualitatively similar performance on the pelvis and sternum object classes. The final instance segmentation is realized by combining the initial vertebral body semantic segmentation with the associated FLT-PET image data, where the vertebral boundaries become well-resolved by the 28th day post-transplant. The vertebral boundary detection algorithm is a hand-crafted spatial filter that enforces vertebra span as an anatomical prior, and it performs similar to a human for the detection of all but one vertebral boundary in the entirety of the HSCT patient dataset. In addition to the segmentation model, I propose, design, and test a “drop-in” replacement up-sampling module that allows state-of-the-art super-resolution convnets to be used for purely asymmetric upscaling tasks (tasks where only one image dimension is scaled while the other is held to unity). While the asymmetric SR convnet I develop falls short of the initial goal, where it was to be used to enhance the unresolved vertebral boundaries of the under-sampled CT image data, it does objectively upscale medical image data more accurately than naïve interpolation methods and may be useful as a pre-processing step for other medical imaging tasks involving anisotropic pixels or voxels. | en_US |
| dc.identifier.uri | https://hdl.handle.net/11244/330194 | |
| dc.language | en | en_US |
| dc.subject | Medical Image Segmentation | en_US |
| dc.subject | Image Super-Resolution | en_US |
| dc.subject | Hematopoietic Stem Cell Transplantation (HSCT) | en_US |
| dc.subject | Convolutional Neural Network (CNN) | en_US |
| dc.thesis.degree | Master of Science | en_US |
| dc.title | Automatic Bone Structure Segmentation of Under-Sampled CT/FLT-PET Volumes for HSCT Patients | en_US |
| ou.group | Gallogly College of Engineering::School of Electrical and Computer Engineering | en_US |
| shareok.nativefileaccess | restricted | en_US |