QUANTITATIVE NUCLEAR MEDICINE IMAGING USING ADVANCED IMAGE RECONSTRUCTION AND RADIOMICS
MetadataShow full item record
Our aim is to help put nuclear medicine at the forefront of quantitation on the path to the realization of personalized medicine. We propose and evaluate (Part I) advanced image reconstruction and (Part II) robust radiomics (large-scale data-oriented study of radiological images). The goal is to attain significantly improved diagnostic, prognostic and treatment-response assessment capabilities. Part I presents a new paradigm in point-spread function (PSF)-modeling, a partial volume correction method in PET imaging where resolution-degrading phenomena are modeled within the reconstruction framework. PSF-modeling improves resolution and enhances contrast, but significantly alters noise properties and induces edge-overshoots. Past efforts involve a dichotomy of PSF vs. no-PSF modeling; by contrast, we focus on a wide-spectrum of PSF models, including under- and over-estimation of the true PSF, for the potential of enhanced quantitation in standardized uptake values (SUVs). We show for the standard range of iterations employed in clinic (not excessive), edge enhancement due to overestimation actually lower SUV bias in small regions, while inter-voxel correlations suppress image roughness and enhance uniformity. An overestimated PSF yields improved contrast and limited edge-overshoot effects at lower iterations, enabling enhanced SUV quantitation. Overall, our framework provides an effective venue for quantitative task-based optimization. Part II proposes robust and reproducible radiomics methods. Radiomics workflows are complex, generating hundreds of features, which can lead to high variability and overfitting, and ultimately hampering performance. We developed and released a Standardized Environment for Radiomics Analysis (SERA) solution to enable robust radiomics analyses. We conduct studies on two unique imaging datasets – renal cell carcinoma SPECT and prostate cancer PET – identifying robust and reproducible radiomic features. In addition, we evaluate a novel hypothesis that radiomic features extracted from clinically normal (non-ischemic) myocardial perfusion SPECT (MPS) can predict coronary artery calcification (CAC; as extracted from CT). This has important implications, since CAC assessment is not commonly-performed nor reimbursed in wide community settings. SERA-derived radiomic features were utilized in a multi-step feature selection framework, followed by the application of machine learning to radiomic features. Our results show the potential to predict CAC from normal MPS, suggesting added usage and value for routine standard MPS.