A 3D-CNN Classifier for Gender Discrimination from Diffusion Tensor Imaging of Human Brain

Research on medical image recognition by machine learning has been studied for a long time. Medical image recognition has been mainly for detection of diseases or symptoms. In recent years, there is a hypothesis which human health conditions and personal attributes affect brain structure. Most studies focus on T1-weighted images, however T2-weighted and diffusion-weighted images are also used in medical practice. Diffusion images contains signals and information which are possibly used to discriminate human health or other state rather and are different from T1-weighted or T2-weighted images. Some studies using 3D-CNN (3-dimensional convolutional neural network) have been applied for brain structural magnetic resonance imaging such as T1-weighted images, but diffusion tensor images cannot be directly input to 3D-CNN because their data is represented as the tensor of the movement of water molecules. Therefore, in this research, we propose 3D-CNN which input the diffusion images of each direction to different channel. We construct multi-channel 3D-CNN and verify the effect of brain structure image mode on gender discrimination using T1, T2, and DTI.