deep-learningLSTMsiamesetractography

Siamese LSTM-based fiber structural similarity network (FS2NET) for rotation invariant brain tractography segmentation

Reviewed on Jun 28, 2018 by Daniel Jörgens • https://arxiv.org/abs/1712.09792.pdf

Reference : Patil, SM and Nigam, A and Bhavsar, A and Chattopadhyay, C. Siamese LSTM based fiber structural similarity network (FS2NET) for rotation invariant brain tractography segmentation. arxiv.org.

Motivation

Streamline classification usually requires registration of subject data to a common space.
Contribution: The main goal of this work is to provide a rotationally invariant approach for streamline segmentation based on a set of reference streamlines.

Main idea

Learn the probability of two streamlines to belong to the same bundle.
Recognise a streamline by assigning the label of the fiber with the highest similarity in a given set of reference fibers.

Rotation invariance

Assumption: Similarity between fibers is learned. This is supposed to generalise to

a. misaligned

b. rotated

streamlines.
Therefore, applying data augmentation to the reference set is sufficient for achieving rotation invariance. (I.e. augmentation independent of the training stage.)

Method

FS2Net:
- stacked LSTMs to arrive at a 32-entry feature vector
- “L1-distance” between feature vectors
  
  Comment: Probably, the authors mean the difference vector.
- classification network: ‘similar’ vs ‘dissimilar’

Experiments

Comparison to two of their own works: ANN and BrainSegNet
- Comment: Doubtful if this can be considered as ‘state of the art’.
Training:
- BS: 11, 1000 batches per epoch -> only 11K training samples. # epochs: not clear.

Nomenclature

macro classification - Gray matter fibres vs White matter fibres
- probably: U-shaped or short fibres vs long range fibres (?)

micro classification - Labelling WM fibres into 8 subclasses
Recall - fraction of correct WM labels in macro classification

Data

3 subjects, 250K streamlines each
preprocessing: keep only 75% of streamline points with the highest curvature (per fibre)
each streamline: max of 100 points (zero padded)

Registered data

FS2Net achieves top performance compared to ANN and BrainSegNet in all experiments.
FS2Net shows better recall hinting at a better robustness towards data imbalance of GM and WM classes (macro).

Unregistered data

FS2Net is quite robust to rotations of testing streamlines.

Comment

Very nice idea. However, for the real value of their approach for streamline segmentation the evaluation is of limited power.