Validating the Utility of Supervised Clustering Algorithm for Precise [11C]DPA-713 PET Brain Image Quantification

TL;DR

This study validates a new supervised clustering algorithm for brain PET imaging, showing it can replace traditional methods and improve accuracy in measuring brain activity.

Contribution

The study introduces and validates a supervised clustering algorithm (SVCA) as a reliable alternative to arterial input function for PET quantification.

Findings

SVCA-DVR showed strong correlation with AIF-DVR, with correlation coefficients of 0.86 and 0.95 in white matter and thalamus, respectively.

Test-retest variability was significantly reduced for SVCA-DVR compared to AIF-DVR across different brain regions and VOI sizes.

SVCA-DVR demonstrated consistent performance even in small volumes of interest, maintaining variability below 5%.

Abstract

The reliance of quantitative PET imaging on the arterial input function makes brain PET challenging to perform in certain populations, limiting the sample size. To address this challenge, a supervised clustering algorithm (SVCA) has been introduced as an alternative. Our objective was to validate SVCA’s performance for brain PET with [11C]DPA-713 that targets a putative marker of brain injury and repair. Methods: This study included a composite dataset comprising 12 healthy volunteers (HVs), with 6 participants from Weill Cornell Medicine and 6 participants from Johns Hopkins University School of Medicine. The minimum number of subjects required to define kinetic classes was identified. Next, the distribution volume ratio (DVR) was examined by comparing pseudoreference time–activity curves derived from SVCA (SVCA-DVR) with the conventional arterial input function–based DVR (AIF-DVR). Test–retest analysis was conducted to evaluate repeatability, considering volumes of interest (VOIs) of various sizes. Lastly, the research investigated differences in DVR values between the HVs and patients with multiple sclerosis. Results: The number of subjects necessary for the kinetic classes, which are critical to SVCA, was reduced to 7 from the existing minimum requirement of 10. This allowed for a more substantial independent validation within a defined dataset. Correlative analysis between SVCA-DVR and AIF-DVR demonstrated a strong relationship, with correlation coefficients of 0.86 for white matter and 0.95 for the thalamus. Furthermore, we noted a marked decline in absolute test–retest variability for SVCA-DVR, with reductions from 1.31% to 1.18% in white matter and 3.51% to 2.32% in the thalamus, relative to AIF-DVR. This pattern of reduced variability persisted across VOIs of disparate sizes, with the absolute test–retest variability remaining below 5% for SVCA-DVR, even in small VOIs (both high and low binding at 0.065 cm3). Analysis revealed a pronounced disparity in SVCA-DVR values of the thalamus when comparing HVs and patients with multiple sclerosis. Conclusion: The findings substantiate the pseudoreference time–activity curves derived from SVCA as a dependable and practical substitute for the quantification of [11C]DPA-713 PET scans of the brain.

Figures