CNS: CycleGAN-Assisted Neonatal Segmentation Model for Cross-Datasets

Accurate segmentation of neonatal brain MR images is critical for studying early brain development. Recently, supervised learning-based methods, i.e., using convolutional neural networks (CNNs), have been successfully applied to infant brain segmentation. Although these CNN-based methods have achieved reasonable segmentation results on the testing subjects acquired with similar imaging protocol as the training subjects, they are typically not able to produce reasonable results for the testing subjects acquired with different imaging protocols. To address this practical issue, in this paper, we propose leveraging a cycle-consistent generative adversarial network (CycleGAN) to transfer each testing image (of a new dataset/cross-dataset) into the domain of training data, thus obtaining the transferred testing image with similar intensity appearance as the training images. Then, a densely-connected U-Net based segmentation model, which has been trained on the training data, can be utilized to robustly segment each transferred testing image. Experimental results demonstrate the superior performance of our proposed method, over existing methods, on segmenting cross-dataset of neonatal brain MR images.