Spherical Harmonic Residual Network for Diffusion Signal Harmonization
Summary
Goal: Harmonize the dMRI signal differences between scanners
Method: Use a ResNet on 3x3x3 patches to harmonize SH coefficients voxelwise, by learning a mapping between sites.
Highlights:
 No need for registration (at test time)
 Modelfree (spherical harmonics do not impose a model on the data)
Model
 RISH projection: At test time, RISH features are used to project the harmonized signal and correct for changes in fiber orientation
where \(S_i'\) are the hamonized SH coefficients and \(S_i\) the nonharmonized SH coefficients of order \(i\).
Reminder:
RISH: Squared L2 norm for each order of SH coefficients
Results
Datasets:
CDMRI harmonization Challenge
 10 healthy subjects, scanned on a 3T GE ExciteHD and a 3T Siemens Prisma, with 30 directions at b=1200.
Training:
 FAST registration for training subjects
 5 epochs using Adam with batch size 256, then SGD with batch size 128 and learning rate decay…
Experiments:
 Investigate the sensitivity of the method to the number of ResBlocks n (Table 1)
 Evaluate the ability to reduce interscanner variance in the signal itself and metrics like FA/MD (Figures 2,3,4)
 10fold crossvalidation using the 10 subjects (8 training, 1 validation, 1 test).
The only reported baseline is the Golkov method^{1}, a 3layer, 150 units neural network with ReLU and Dropout designed, applied on a single voxel at a time; it was “designed” for the estimation of scalar measures from diffusion data, and seems completely unrelated to harmonization.
There is no classical method baseline…

Golkov, V., Dosovitskiy, A., Sperl, J.I., Menzel, M.I., Czisch, M., Sämann, P., Brox, T. and Cremers, D., 2016. Qspace deep learning: twelvefold shorter and modelfree diffusion MRI scans. IEEE transactions on medical imaging, 35(5), pp.13441351. ↩