Code: []


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


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.



  • 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  • May be considered outdated compared to other methods, e.g. CLARITY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


  • 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


  • 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


  • 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.