TODO.md

#Data preprocessing

• dwi values are int16, so it should be normalized, either by:

  • using the b0 <-
    • do we just divide by the b0 (element-wise)?
  • using the int16.max()

• if using original hcp data, we should perform a top-up

• streamlines

  • resample streamlines so they have 100 points each.

#LSTM - regression

Inputs

Diffusion weights in all available direction

Outputs

1. Direction to follow

• as-is
• normalized

2. Continue or stop (using the binary cross-entropy)

• Since it will be higly unbalanced, when the target is "stop" multiply the cost by the number of "continue" for a given streamline.

#LSTM - classification

Inputs

Diffusion weights in all available direction

Outputs

1. Softmax of the direction (something about softmax of gaussians)
2. Continue or not (using the binary cross-entropy)

• Since it will be higly unbalanced, when the target is "stop" multiply the cost by the number of "continue" for a given streamline.