A Knowledge-Based Artificial Intelligence (AI) to Perform Nested Model Selection From Dynamic Contrast Enhanced (DCE)-MRI Pharmacokinetic Analyses of Brain Tumors in An Animal Model
H Bagher-Ebadian1,2, T Nagaraja1, G Cabral1, K Farmer1, O Valadie1, P Acharya2, B Movsas1, S Brown1, I Chetty1*, J Ewing1, (1) Henry Ford Health System, Detroit, MI, (2) Oakland University. Rochester, MI
Presentations
WE-C930-IePD-F6-2 (Wednesday, 7/13/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 6
Purpose: Our group has shown that a nested model selection (NMS) technique utilizing an extended Patlak graphical method illuminates pharmacokinetic (PK) compartmental analyses of DCE-MRI data. NMS generates maps of brain regions that reflect the number of parameters needed to describe their vascular physiology based on DCE-MRI. However, identification of model choice regions requires a series of computationally intensive processing. Furthermore, prediction of model-1 region is biased by dispersion of the arterial input function (AIF). In this study, we introduce a knowledge-based adaptive model for real time prediction of MS maps from DCE-MRI raw information while minimizing AIF dispersion error.
Methods: Thirty-nine immune compromised RNU rats were implanted with human U-251n cancer cells orthotopic glioma (IACUC: #1509). Sixty-six DCE-MRI studies (28 days after tumor implantation, 7T-Dual-Gradient-Echo, FOV:32x32mm2, TR/(TE1-TE2)=24ms/(2ms-4ms), flip-angle=18º, 400-acquisitions/1.55sec-interval, Magnevist/tail-vein) were used to perform PK analysis using NMS to distinguish three different brain regions: Normal vasculature (Model-1: No-leakage), leaky tumor tissues with no back-flux to vasculature (Model-2), and leaky tumor tissues with back-flux (Model-3). Normalized time-traces of DCE-MRI (1st-and 2nd-echoes) for three model regions (92698 profiles/examples) were used to train (target: PK-MS results) and validate (10-fold nested cross validation, NCV) an interconnected or nested artificial neural networks (ANNs with feed forward architecture: 76:10:1, Levenberg-Marquardt optimization, loss: cross-entropy). To suppress the AIF-dispersion errors, a knowledge-based optimization was performed on the ANNs’ response distributions.
Results: The NCV performance (outer-loop) of the trained-ANNs for the prediction of Model-1 versus Models-2 & 3, and Model-2 versus Model-3 were: AUC/F1-Score/Balanced-Accuracy= 0.909/0.833/0.850 and 0.955/0.884/0.890, respectively. Compared to the conventional NMS analysis, the Model-1 regions predicted by the ANNs were less impacted by the AIF-dispersion effects (less miss-classification for Model-1 and 2).
Conclusion: This pilot study demonstrates the use of adaptive models for PK analyses of DCE-MRI data that characterizes the vascular physiology of embedded tumors.
Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by a grant from Varian Medical Systems (Palo Alto, CA) and NCI/NIH R01-CA218596.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email