We investigate the problem of infection diagnosis in hydrocephalic images. While traditional deep learning methods can be used for this problem - they require generous training data and produce class activations that are not meaningful. A novel 2D brain attention regularized network (BAR-Net) is proposed in our work, which encourages the network to focus inside the brain region for classification decisions. The newly proposed regularizer yields more interpretable class activations and experimental results confirm that the issue of over-fitting is mitigated. Extending the 2D BAR-Net, a hybrid 2-D/3-D CNN is proposed, that facilitates interactions between the 2D and 3D branches. Furthermore, a mutual attention regularization term is introduced in order to enable branches to exchange knowledge learned from the other by transferring attention maps for regularization.