Investigating Whether the Addition of Advanced Quantitative MRI Improves Prediction Accuracy of MS White Matter Lesion Evolution Compared to Conventional MRI In a 3D U-Net
Abstract
Purpose
Previously, we showed equivalent success in predicting multiple sclerosis (MS) white matter lesion evolution over two years using separate 3D U-Nets based on either baseline conventional MRI (cMRI) or quantitative MRI (qMRI). Here, we investigated 3D U-Net model performance using both separate and combined cMRI and qMRI inputs.
Methods
Baseline MRI data, binary white matter (WM) and white matter lesion (WML) masks from 35 persons with MS were pre-processed and nonlinearly normalized to the MNI152 template space. Baseline and two-year follow-up MWL masks were used to compute a 4-class (i.e., shrinking, stable, growing, and background) disease evolution map (DEM) for each participant. Three sets of channel-wise 4D-concatenated inputs were then used to train separate 3D U-Nets to minimize the 4-class DEM cross-entropy, using participants’: 1) baseline WM masks and ten different WM-cropped qMRI maps, 2) baseline WM masks and two WM-cropped and intensity-normalized cMRI images, or 3) baseline WM masks and all available MRI maps. Due to computational resource limitations, these 4D volumes were subsampled into smaller 4D blocks (32 x 32 x 32 x channel number). Each model was trained using leave-one-out cross-validation (n=35). We evaluated each model (OverallqMRI, OverallcMRI, Overallcombined) using Dice similarity coefficients (DSCs) between the predicted and ground truth follow-up DEMs. Overallcombined model performance was compared to OverallqMRI and OverallcMRI using two-tailed, Wilcoxon signed-rank tests.
Results
DSC values for the combined model [median (IQR) = 0.73 (0.66 – 0.80)] did not differ compared to the cMRI model [0.73 (0.64 – 0.79)] or the qMRI model [0.72 (0.60 – 0.80)] after correcting for multiple comparisons (pFWE > 0.05).
Conclusion
While qMRI methods have clinical applications in MS and may improve MS lesion evolution prediction using other deep learning methods, their inclusion in our 3D U-Net did not significantly improve model performance over the use of cMRI.