Figure 2
Scheme of the Proposed Solution. (a) On the left-hand side, a multiparametric representation of the imaging is created via fusion of corresponding slices from different sequences into the colour channels of the RGB model. In the centre, the label map is first divided in tumour region (RT) and the background regions (RB) according to the delineation done by the experienced reader. On the right-hand side, N voxels (together with their surrounding patch) are then randomly sampled from these regions to maintain a balance between number of voxels representing the tumour and number of voxels representing healthy tissue. (b) The architecture of the network, which is trained with the patches of the images in the discovery set. The patches of the images in the test set are used to control for model overfitting. (c) The 3D probability map is generated by classification of each voxel using the trained model. The probability map is thresholded to find the components where the probability of tumour is higher than the probability of healthy tissue. The largest component is selected as segmentation of the tumour.
