Fig. 2.

Overview of the proposed deep learning architecture. Top left: The network takes five MRI images (2D slices from DWI, ADC, CBV, CBF, T_max volumes) as input. Below: Each input image is processed independently on 5 separate branches. Pink, purple, yellow, red and green feature maps result from 2D-convolutions and maxpooling. The output of the 5 branches are then concatenated, and upsampled through 2D-deconvolution layers. The network produces an output map with 3 classes (lesion, healthy tissue and background). Top Right: The predicted lesion has to be compared to the true lesion from the final FLAIR.