Precise 3D segmentation of infant brain tissues is an essential step towards\ncomprehensive volumetric studies and quantitative analysis of early brain\ndevelopement. However, computing such segmentations is very challenging,\nespecially for 6-month infant brain, due to the poor image quality, among other\ndifficulties inherent to infant brain MRI, e.g., the isointense contrast\nbetween white and gray matter and the severe partial volume effect due to small\nbrain sizes. This study investigates the problem with an ensemble of semi-dense\nfully convolutional neural networks (CNNs), which employs T1-weighted and\nT2-weighted MR images as input. We demonstrate that the ensemble agreement is\nhighly correlated with the segmentation errors. Therefore, our method provides\nmeasures that can guide local user corrections. To the best of our knowledge,\nthis work is the first ensemble of 3D CNNs for suggesting annotations within\nimages. Furthermore, inspired by the very recent success of dense networks, we\npropose a novel architecture, SemiDenseNet, which connects all convolutional\nlayers directly to the end of the network. Our architecture allows the\nefficient propagation of gradients during training, while limiting the number\nof parameters, requiring one order of magnitude less parameters than popular\nmedical image segmentation networks such as 3D U-Net. Another contribution of\nour work is the study of the impact that early or late fusions of multiple\nimage modalities might have on the performances of deep architectures. We\nreport evaluations of our method on the public data of the MICCAI iSEG-2017\nChallenge on 6-month infant brain MRI segmentation, and show very competitive\nresults among 21 teams, ranking first or second in most metrics.\n