Multi-objective parameter optimization of DWI-based global fiber tracking with neuronal tracer signal as a reference

Diffusion magnetic resonance imaging (MRI) allows non-invasive exploration of brain structural information, crucial for understanding the human-brain and related neurological and mental diseases; however, present-day diffusion-based tractography approaches are still lacking in accuracy. Improvement efforts included comparison with anatomical or structural connectivity data, the usage of higher quality or complementary data, more sophisticated fiber tracking algorithms and parameters assessment. Despite that, well-known problems such as low sensitivity, dominance of false positives, the accurate reconstruction of long-range connections could not be avoided or minimized. Current methods include several parameters, typically tuned in a heuristic way, with different settings to increase the sensitivity of different pathways, without attenuating the overall problem. We propose a new way to optimize fiber-tracking results by multi-objectives tailored to minimize the current issues. Our optimization approach uses comparison with marmoset fluorescent tracer injection data from the Japan Brain/MINDS project to validate and optimize ex-vivo diffusion weighted MRI (DWI)-based global fiber tracking results of the same subject. The approach includes multiple comparisons at a”higher-than-DWI” resolution standard brain space to evaluate objectives of coverage and passage constraint. The coverage objectives maximize sensitivity while constraining the growth of false positives, promoting the generation of long connections to the projection areas. The passage constraint minimizes the number of fibers crossing hemispheres outside the commissures. We implemented a non-dominated sorting genetic algorithm II (NSGA-II) strategy to explore the parameters space and evaluate our multi-objectives through evolution. Evolutionary optimization runs for 10 brain samples in parallel while sharing”champion” settings over samples, encouraging convergence of parameters in a similar locus. We evaluate the generalization capability of the optimized parameters on unseen marmoset test data leading to promising results.