Automatic Vertebral Body SUV Extraction of Low Dose FLT-PET/CT Scans for HSCT Patients

Abstract

In this thesis, I present a novel, fully automated framework for multi-class vertebral body segmentation and standardized uptake value (SUV) extraction in low-dose FLT-PET/CT scans of hematopoietic stem cell transplantation (HSCT) patients. The approach combines a custom-made, attention-gated 3D U-Net with label augmentation strategies derived from publicly available datasets, specifically designed to address the challenges of coarse axial CT resolution and limited annotated clinical data.Quantitative evaluations using the Dice similarity coefficient and SUV error analysis, along with qualitative assessments via 3D renderings and difference maps, demonstrate high concordance between automated SUV measurements and physician-derived ground truths—even in anatomically variable regions and under modest segmentation accuracy. The model achieved an average Dice score of 0.83 across all validation cases and a mean SUV correlation exceeding 0.95. Unlike previous methods, this pipeline reliably segments spines exhibiting anomalies such as a lumbarized S1 vertebra, which tend to 'confuse' existing networks and result in underestimation of vertebral count. The pipeline offers a clinically actionable approach for identifying and monitoring metabolic changes in vertebral marrow cavities, facilitating earlier detection of engraftment success or potential relapse. In contrast to existing multi-step or partially automated workflows, this end-to-end solution demonstrates stable performance across diverse anatomical conditions. To the best of my knowledge, this is the first work to directly segment all 24 individual vertebral bodies in CT images with 5 mm axial slice spacing. The results underscore the feasibility of extracting accurate SUV data from under-sampled FLT-PET/CT scans and provide a foundation for future research into robust, patient-specific relapse prediction models.