Going deeper with convolutions
We present our entry for the Longitudinal Multiple Sclerosis Challenge 2015 using 3D convolutional neural networks (CNN). We model a voxel-wise classifier using multi-channel 3D patches of MRI volumes as input. For each ground truth, a CNN is trained and the final segmentation is obtained by combining the probability outputs of these CNNs. Efficient training is achieved by using sub-sampling methods and sparse convolutions. We obtain accurate results with dice scores comparable to the inter-rater variability.