Code: [https://github.com/finalelement/dl_untapped_info_dwmri_histo]

Highlights

• ResDNN trained to map a raw diffusion MRI signal to a fODF
• Ground truth built from histology
• Outperforms CSD reconstruction on histology measurement

Method

Input: 45 SH coefficients (order 8) fitted to ex-vivo DW-MRI Target: Histological fiber ODFs, represented by 45 SH coefficients (order 8)

Model: MLP with a residual block, with architecture: [45-400-45-200-45-200-45]

No validation set was used; cross-validation is used to determine to “optimal” number of iterations for convergence. Training time is less than 10min and prediction is done in less than 2min.

Experiments

Data

• 3 Squirrel monkey brains (2 for training, 1 for testing)
• HCP scan-rescan of 12 subjects

Ground truth for Squirrel monkey

• Computed using slice-wise staining and 3D structure tensor analysis using a confocal microscope
• Histology registered to DW-MRI data
• May be considered outdated compared to other methods, e.g. CLARITY
• Total of around 50000 voxels, including rotations as data augmentation

Baselines

• Q-Ball Imaging
• Q-Ball Imaging with constant solid angle
• Super-resolved CSD at SH order 6 and 8
• Lucy-Richardson CSD
• Diffusion orientation transform
• Diffusion orientation transform revisited

Results

• Evaluation uses ACC (Angular correlation coefficient), a measure of similarity for SH coefficients, ranging from -1 to 1

Results on Squirrel monkey

• ResDNN median: 0.82
• All others: <0.79

Results on HCP 12 subjects scan-rescan

• ResDNN: 0.74 +- 0.31
• CSD: 0.61 +- 0.31

Comments

• The model processes SH coefficients fitted to the log-space signal, and uses ReLU activations. These 2 details enforce non-negativity in SH space, as seen in Figure 8.