Optimizing Rapid Diffusion Kurtosis Imaging Sequences for Clinical Translation In Radiation Oncology
Abstract
Purpose
To optimize diffusion-kurtosis imaging (DKI) MRI acquisitions for improved time efficiency while preserving quantitative data fidelity, enabling the future integration of DKI into cancer-care workflows for RT treatment planning and disease tracking.
Methods
A pineapple was chosen as a diffusion phantom for its large size and varied structure. It was imaged at 15 non-zero b-values commonly used in quantitative diffusion methods (30-4000 s/mm2). 8 images were averaged per b-value and gradient direction to create the set of 15 diffusion weighted images. Mean diffusivity and kurtosis were derived using in-house code as baseline estimates from the full data set. Using an ablation study approach, a greedy search algorithm was developed to continuously reduce the set of b-values for parameter fitting. Hotelling’s T2 was computed as an error metric for comparisons to the baseline at each algorithm iteration. A reduced b-value subset maintaining high parameter accuracy was chosen by applying the elbow method to the T2 plot.
Results
Following the ablation procedure, the T2 values were plotted as a function of the number of b-values removed from the fits. The elbow point was identified at 10 removed b-values, with an optimized subset of 5 non-zero b-values (300, 500, 1500, 2500, 4000 s/mm2). The T2 comparing the fits from this subset to the baseline was 7.4, and beyond being merely an elbow point, it corresponded to a local minimum in the plot.
Conclusion
This work displays the feasibility of incorporating DKI into cancer-care workflows, cutting scan times by over half. A novel greedy search algorithm successfully reduced the number of b-values required to perform accurate fits of diffusivity and kurtosis. Future work involving phantom and volunteer scans will incorporate perfusion parameters into the diffusion signal equation. The algorithm will also be expanded to optimize for the number of images averaged per b-value.