Investigating the Discrimination Ability of 3D Convolutional Neural Networks Applied to Altered Brain MRI Parametric Maps

Convolutional neural networks are gradually being recognized in the neuroimaging community as a powerful tool for image analysis. In the present study, we tested the ability of 3D convolutional neural networks to discriminate between whole-brain parametric maps obtained from diffusion-weighted magnetic resonance imaging. Original parametric maps were subjected to intensity-based region-specific alterations, to create altered maps. To analyze how position, size and intensity of altered regions affected the networks’ learning process, we generated monoregion and biregion maps by systematically modifying the size and intensity of one or two brain regions in each image. We assessed network performance over a range of intensity increases and combinations of maps, carrying out 10-fold cross-validation and using a hold-out set for testing. We then tested the networks trained with monoregion images on the corresponding biregion images and vice versa. Results showed an inversely proportional link between size and intensity for the monoregion networks, in that the larger the region, the smaller the increase in intensity needed to achieve good performances. Accuracy was better for biregion networks than for their monoregion counterparts, showing that altering more than one region in the brain can improve discrimination. Monoregion networks correctly detected their target region in biregion maps, whereas biregion networks could only detect one of the two target regions at most. Biregion networks therefore learned a more complex pattern that was absent from the monoregion images. This deep learning approach could be tailored to explore the behavior of other convolutional neural networks for other regions of interest.