Frequency-dependent diffusion tensor distribution imaging in the evaluation of ischemic stroke

Non-invasive magnetic resonance imaging (MRI) is considered the gold standard for the prognosis and monitoring of ischemic stroke. Still, the conventional methods used in clinics often fail to detect subtle changes and estimate tissue viability in the lesion, penumbra and distal regions. In this study, we combined frequency-dependent distribution tensor imaging (ωDTD) and clustering of the frequency dependent diffusion tensor distributions D(ω) with multivariate regression modelling in a whole brain section to interpret ischemic changes in quantitative tissue microstructural parameters. We performed ex vivo ωDTD and histology in middle cerebral artery occlusion (MCAO) and sham-operated rats (P = 17) 24 hours post reperfusion. The lesions were characterised by cell loss and a presence of smaller cells, most likely glial cells. A random forest (RF) model was used to explain and predict the histological parameters based on diffusion tensor imaging (DTI), manually bin-resolved and cluster-resolved ωDTD parameters. Cross-validation with leave-one-animal-out (CV LOO) was used to evaluate our model. We found that ωDTD parameters were more representative of the number of cells compared to diffusion tensor imaging (DTI) metrics (ωDTD R^2 = 0.73 vs DTI R^2 = 0.49 with CV ωDTD R = 0.23 vs DTI R = 0.33), area and circularity of nucleus (ωDTD R^2 = 0.64 vs DTI R^2 = 0.40 and ωDTD R^2 = 0.61 vs DTI R^2 = 0.35 with CV ωDTD R = 0.13 vs DTI R = 0.26 and ωDTD R = 0.17 vs DTI R = 0.31) than DTI. We also found that a more flexible modelling approach, such as RF, was advantageous to represent the complex, non-linear relationship between MRI and tissue. In conclusion, this study shows that ωDTD, along with advanced machine learning methods, has the potential to help improve the understanding and interpretation of ischemic stroke related tissue damage in terms of advanced MRI.