Abstract:Purpose: Deformable image registration is a fundamental task in medical imaging. Due to the large computational complexity of deformable registration of volumetric images, conventional iterative methods usually face the tradeoff between the registration accuracy and the computation time in practice. In order to boost the performance of deformable registration in both accuracy and runtime, we propose a fast unsupervised convolutional neural network for deformable image registration. Methods: The proposed registration model FDRN possesses a compact encoder-decoder network architecture which employs a pair of fixed and moving images as input and outputs a three-dimensional displacement vector field (DVF) describing the offsets between the corresponding voxels in the fixed and moving images. In order to efficiently utilize the memory resources and enlarge the model capacity, we adopt additive forwarding instead of channel concatenation and deepen the network in each encoder and decoder stage. To facilitate the learning efficiency, we leverage skip connection within the encoder and decoder stages to enable residual learning and employ an auxiliary loss at the bottom layer with lowest resolution to involve deep supervision. Particularly, the low-resolution auxiliary loss is weighted by an exponentially decayed parameter during the training phase. In conjunction with the main loss in high-resolution grid, a coarse-to-fine learning strategy is achieved. Last but not least, we involve a proposed multi-label segmentation loss (SL) to improve the network performance in Dice score in case the segmentation prior is available. Comparing to the SL using average Dice score, the proposed SL does not require additional memory in the training phase and improves the registration accuracy efficiently. Results: We evaluated FDRN on multiple brain MRI datasets from different aspects including registration accuracy, model generalizability, and model analysis. Experimental results demonstrate that FDRN performs better than the state-of-the-art registration method VoxelMorph by 1.46% in Dice score in LPBA40. In addition to LPBA40, FDRN obtains the best Dice and NCC among all the investigated methods in the unseen MRI datasets including CUMC12, MGH10, ABIDE, and ADNI by a large margin. Conclusions: The proposed FDRN provides better performance than the existing state-of-the-art registration methods for brain MR images by resorting to the compact autoencoder structure and efficient learning. Additionally, FDRN is a generalized framework for image registration which is not confined to a particular type of medical images or anatomy.

FDRN: A fast deformable registration network for medical images