Abstract
The development of quantitative-imaging (QI)-based biomarkers to guide clinical decision-making, such as identifying patients with vs. without disease, is of strong interest across nuclear medicine applications. To address the lack of ground truth in clinical settings, no-gold-standard evaluation (NGSE) approaches such as regression-without-truth (RWT) have been developed. However, existing NGSE approaches primarily assess measurement precision rather than the ability to stratify patient populations, partly because they assume that the underlying true values follow a unimodal distribution, whereas effective biomarkers for binary patient stratification are expected to follow a bimodal distribution. To address this gap, we developed bimodal-RWT (BM-RWT), an NGSE technique that evaluates QI methods based on their population-stratification ability.
BM-RWT models the true quantitative values using a truncated bimodal normal distribution while retaining the linear measurement model of prior NGSE techniques. A maximum-likelihood estimator is derived to estimate model parameters without access to ground truth. The noise-to-slope ratio (NSR) is used as the figure of merit, where a smaller NSR corresponds to better ability to discriminate patient populations. BM-RWT was validated in numerical experiments and in an in silico imaging trial evaluating three attenuation compensation (AC) methods for dopamine transporter single-photon emission computed tomography (DaT-SPECT) on the task of distinguishing patients with normal vs. reduced striatal binding ratio (SBR). Three AC methods were evaluated: CT-based attenuation correction (CTAC), uniform attenuation correction (UAC), and a deep-learning CT-less approach (CTLESS).
In numerical experiments with three hypothetical methods evaluated across 200 independent noise realizations, BM-RWT recovered the correct NSR-based method ranking in 172 out of 200 realizations, with accurate estimation of noise parameters and reasonable recovery of the latent bimodal distribution parameters. The NSR ranking also reflected each method's ability to separate the two latent subpopulations, demonstrating that BM-RWT provides a richer characterization of method performance than classical unimodal RWT. In the in silico imaging trial, BM-RWT yielded the same method ranking as ground-truth-based evaluation, correctly identifying CTAC and CTLESS as outperforming UAC. These results demonstrate that BM-RWT can reliably rank QI methods for patient population stratification without ground truth, motivating further validation across other clinical applications. In the in silico imaging trial, BM-RWT yielded the same method ranking as ground-truth-based evaluation, correctly identifying CTAC and CTLESS as outperforming UAC. These results demonstrate that BM-RWT can reliably rank QI methods for patient population stratification without ground truth, motivating further validation across other imaging-biomarker-based clinical applications.
Committee Chair
Abhinav K. Jha
Committee Members
Scott Norris,Vladimir Kurenok
Degree
Master of Science (MS)
Author's Department
Electrical & Systems Engineering
Document Type
Thesis
Date of Award
Spring 5-6-2026
Language
English (en)
Recommended Citation
Cheng, Qian, "A Bimodal Regression-Without-Truth Technique for Objective Evaluation of Quantitative Nuclear-Medicine Imaging Methods" (2026). McKelvey School of Engineering Theses & Dissertations. 1340.
https://openscholarship.wustl.edu/eng_etds/1340