A Novel Algorithm to Produce Population-Based Input Function for Arterial Input Function

1991 

Abstract Submission 

Daren Ma¹, Sneha Pandya², Ashish Raj¹, Ajay Gupta³

¹University of California San Francisco, San Francisco, CA, ²Weil Cornell Medicine, New York, NY, ³Weil Cornell Medicine, New York , NY

Daren Ma  
University of California San Francisco
San Francisco, CA

Sneha Pandya  
Weil Cornell Medicine
New York, NY

Ashish Raj  
University of California San Francisco
San Francisco, CA

Ajay Gupta  
Weil Cornell Medicine
New York , NY

The [11-C]DPA-713 is a second-generation TSPO ligand and a putative PET imaging biomarker for neuroinflammation that has shown promise in preclinical and clinical studies. Accurate quantification of tracer uptake and metabolism in the tissue through kinetic modeling is crucial but requires the invasive measurement of the tracer concentration in the arterial blood over time. Non-invasive alternatives, like IDIF and population-based input function (PBIF) have gained attention, but their accuracy still relies on additional arterial or venous samples, undermining the non-invasive advantages. We propose a completely non-invasive approach that uses the PBIF technique but does not require calibration samples of any kind, utilizing machine learning to learn the input function calibration factor from very limited and easily-available clinical-demographic data. We propose a further ML step to improve the resulting kinetic modeling outcomes, in this case the regional Logan volume of distribution.

The workflow begins with the generation of Time-Activity (TAC) Curves. These curves undergo the Peak Alignment process to derive the AIF, which is normalized by area under the curve (AUC) of true AIF. The PBIF is obtained as the mean of normalized time activity curves from all subjects. The 1st ML Model is a Gradient Boosting Regressor, which serves as an AUC Predictor, using its output to de-normalize PBIF and estimate subject-specific pAIF. Kinetic modelling is then done to derive Logan V_T from both empirical and predicted AIF for each region. Following which, Logan V_Ts from pAIF are validated against V_T of the same regions using empirical AIF through Pearson correlation. The workflow concludes with the 2nd ML Model which estimates the slope of the V_T-V_T relationship. The predicted slope is a scalar per subject utilized to improve the fit between V_Ts from empirical and predicted AIF.

The Results of our two ML models are displayed in the uploaded Figures.
The first figure shows the computational steps to get our PBIF curve as well as the predicted AIFs. A shows the raw TACs before processing. B shows the TAC curves after linear interpolation and peak alignments. We divided TACs in B by their corresponding AUCs, and averaged those to get the PBIF curve (C). D shows the AUC predictions from our Gradient Boosting model. Then in E, we took the PBIF and multiplied by the subject-wise predicted AUCs to form our predicted AIF curves. F displays the rooted mean-square error results accompanied with the average of pAIF lines.
The second figure illustrated the results of our Slope Re-scaling method.
Subplot A shows the scatter plot of Logan V_T computed by kinetic analysis using subject-specific AIFs against those computed using input function obtained from a naive population-based averaging. B: Scatter plot of gold standard Logan V_T against that obtained using our PBIF. The Pearson correlation between the two methods give R² = 0.952, which is significantly higher than the naive method in Panel A. C: The Logan V_T predictions after slope re-scaling. In this plot, most of the data points align with the diagonal line, giving an even higher R² = 0.986. Panels D, E, F show the effect of subjects' diagnosis, sex, and genotype, respectively, reporting their Pearson's R on the side. There is no significant difference in the predictions, suggesting our results are not biased by these groupings. Panels G, H, I show the distribution of predicted Logan V_T corresponding to those categories respectively.

On both tasks we report excellent accuracy, as tested on a cohort of healthy subjects and patients with Parkinson's disease. This study may pave the way for future utility of quantitative microglial imaging using [11-C]DPA-713 in a data-sparse environment without the need for any blood sampling.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ²

Classification and Predictive Modeling

PET Modeling and Analysis ¹

PET

ADULTS

Degenerative Disease

Machine Learning

Positron Emission Tomography (PET)

Other - Parkinson's Disease