Multi-atlas segmentation (MAS) is currently one of the major approaches for medical image segmentation. This approach usually consists of two steps, multi-atlas registration (MAR) and label fusion. In MAR, multiple annotated training images are registered to target image to generate “transferred” labels. In label fusion, the transferred labels are combines to compute target labels by means of weighted voting. A popularity of multi-atlas approach comes from the generality of this approach thanks to the recent presence of large scale annotated datasets e.g. ADNI brain MRI or OAI knee joint MRI datasets. However, the MAS on these large scale datasets has often suffered from the registration error. We propose a novel method of MAR for MAS inspired by the template generation [1] and deep neural network. We present preliminary results of our multi-layered atlas registration (MLAR) on brain MRI MAS.Five annotated brain MRI were incorporated from the training data of MRBrainS challenge [2]. These 3D T1-weighted scans were acquired on a 3.0T Philips Achieva system. All scans have voxel size of 1.0mm x 1.0mm x 1.0mm, TR of 7.9ms, and TE of 4.5ms. All scans were manually labeled by experts into 8 labels; cortical gray matter, basal ganglia, white matter, white matter lesions, cerebrospinal fluid, ventricles, cerebellum, and brainstem. We construct the multi-atlas for each image in leave-one-out manner; for one target image, other four images and labels are used as atlas. In conventional multi-atlas registration (CMAR) for MAS, the atlas labels are directly registered to the target image as shown in Fig. 1 (a). It is straightforward however CMAR suffers from the registration error. Recently, a novel method of duplicating atlases [1] by generating additional atlas images and labels is proposed. In [1], the new atlases are generated by registering the given atlases to other unannotated images called “templates.” We extend this idea to propose the single-layered atlas registration (SLAR) framework which generates the new atlases by registering the atlases to other training atlases, as shown in Fig. 1 (b). In the figure, we first perform atlas-by-atlas, so-called intra-atlas, image registration to generate NxN atlases, where N is a number of atlases. These extended set of atlases is then registered to the target image to transfer NxN atlas labels. Furthermore, inspired by the deep neural network structure, we introduce MLAR framework by adding another type of registration layer to the SLAR, as shown in Fig. 1 (c). In a new layer called the label-wise registration layer, the registered atlases through SLAR are rectified according to the ground-truth label for the registration target image SLAR. Since the input atlas of label-wise registration layer is registered label-by-label to the ground-truth label of the previously image-based-registered atlas, the number of atlases does not change in this layer. In MLAR, these rectified set of atlases is then registered to the target image to transfer NxN atlas labels. In segmentation, the globally-weighed voting (GWV) scheme with the atlas similarity weight is used as label fusion. In GWV, the atlas similarity weight is computed by the patch-wise squared distance of registered and target images.Fig. 2 shows example images of segmentation results obtained from (a) CMAR, (b) MLAR, and (c) manual segmentation. As shown in the figure, CMAR often results in registration error which degenerates the segmentation accuracy as in yellow box. MLAR reduces this registration error with the help of intra-atlas registration and label-wise rectification. As a result, the results obtrained from MLAR are the most visually similar to manual segmentation results. To validate the performance, Dice similarity coefficients (DSC) for testing labels are computed, as shown in Fig. 3. In evaluation, the 8 labels are combined and rejected to generate 3 labels; (1) cerebrospinal fluid including ventricles, (2) gray matter consisting of cortical gray matter and basal ganglia, and (3) white matter including white matter lesions. As shown in the figure, MLAR improved the overall DSCs compared to those of SLAR and CMAR. In cerebrospinal fluid, average DSCs for CMAR, SLAR, and MLAR are 68.14%, 70.06%, and 70.19%, respectively. In gray matter, average DSCs for CMAR, SLAR, and MLAR are 74.50%, 76.56%, and 77.04%, respectively. Finally, in white matter, average DSCs for CMAR, SLAR, and MLAR are 80.03%, 82.25%, and 82.93%, respectivelyA novel framework of MAR with multiple layers was produced. In experiments, we showed promising results of our framework in brain MAS. Extensive validation on brain MRI MAS is remained as future works. To propose additional registration layers such as similar atlas selection layer, will be further focused as future works.