Highlights

• Deep learning for dMRI harmonization, leveraging limited datasets using multi-task learning (multiple independent networks are merged to accomplish a final task).

Method

• Task: Predict multiple high quality targets (modern platforms) from a single low-quality input (out-of-date platform)

• Standard methods
  • Registration to a common template
  • Align mean and variance from each platform (voxelwise/regionally/whole image)
• Deep learning methods
  • CNN-RISH (Rotation-Invariant Spherical Harmonics)
  • DIQT (Deeper Image Quality Transfer)
  • SHResNet

All existing deep learning methods use a single CNN to accomplish a single task (mapping the source to a single target platform). Traveling heads datasets (one subject scanned on multiple scanners) are very limited, which is problematic for deep learning.

Solution

• Use multi-task learning to train multiple networks on independent tasks, then use the last layers as input to a new network adated for the specific task.
• The final prediction is a linear combination of the task-specific predictions.

Model figure

Experiments

Data

• MUSHAC dataset
  • 10 healthy subjects scanned on 3 different scanners
  • GE 3T (“out-of-date” platform), Siemens Prisma 3T, Siemens Connectom 3T
  • Standard (st) protocol: 2.4mm, 30 directions; SotA (sa) protocol: 1.2mm, 60 directions
• Model input: SH coefficients after deconvolution, 6th order (28 coefficients), processed as patches of 11^3 voxels
• Target: Patches of 11^3 (harmonization) or 11^9 (harmonization + super-resolution)

Results

Comments

• MSP reuses many pretrained models, and adds many new sub-networks, so it has much more parameters than any single model, which might give it an advantage over others.
• The final MSP layer is still task-specific.
• No subjects were specifically left out for the test set.
• They seem to choose the task-specific models based on the test set evaluation, which may indicate data contamination.