120 — Improving automated multiple sclerosis lesion segmentation with a cascaded 3D convolutional neural network approach

Valverde et al (1702.04869)

Read on 18 December 2017
#multiple-sclerosis  #CNN  #3D  #neural-network  #MRI  #segmentation  #lesion 

One sign of disease progression in multiple sclerosis (MS) patients is the presence of lesions in the brain: These lesions are often indications of further disease progression even before behavioral deficits arise.

Naturally, it is a challenge to scan every brain MRI for evidence of these lesions, and so an automated system is desireable. This paper presents a method for automated image segmentation — essentially, detection of these lesions — that outperforms other segmentation algorithms when run on the MICCAI2008 MS MRI dataset.

The algorithm is a progression of preprocessing that leads to two convolutional 3D neural networks: First, the image is subtracted from its mean and divided from its variance in order to globally normalize the images between scans. Then the image is divided into small 3D patches (11vox³). These patches are used to train CNN1, the output of which is used to train CNN2, which, which uses $FC$ layers to return the voxelwise probability of belonging to a positive “lesion” or negative “non-lesion” class.

This method augments the training data by flipping along the axial plane to generate more training 3D patches. The authors posit that this does not affect the validity of the training data because at the scale of these patches, the reflection does not dramatically perturb the neuroanatomy contained in the image.

This algorithm outperforms all others on the leaderboard of the MICCAI 2008 MS MRI leaderboard for segmenting white matter lesions.