Abstract
The segmentation of pediatric brain MRI into distinct tissues is important for the evaluation of pediatric brain development and the diagnosis of neurological and neurodevelopmental disorders. However, when the used dataset diverges due to various acquisition protocols or biases among patient cohorts, existing deep learning algorithms cannot guarantee correct predictions. Unsupervised domain adaptation approaches have lately shown enormous potential for addressing this problem by limiting the divergence between the distributions of the used datasets. In this paper, we firstly developed a model called 3DUDRSeg, a 3D encoder-decoder for precise autonomous segmentation of pediatric brain tissues. Our proposed 3DUDRSeg model achieved a 98.88% DSC accuracy rate because of the employment of denes blocks and residual units that help relieve the degradation problem during training the network, allowing the performance advantages to be fully utilized. With this approach, our 3DUDRSeg can create more strong features to deal with the wide range of brain tissue variations. Then, we present 3DAdGanSeg, an entropy-based unsupervised domain adaptation framework for segmenting pediatric brain tissues in unannotated datasets via adversarial learning. The suggested model significantly influences the capability to distinguish the borders between tissue classes, with DSC of 85% and HD95 of 1.479 in the case of the dHCP dataset as the source domain and DSC of 81 % and HD95 of 2.061 when using the Schizophrenia Bulletin 2008 dataset as a source domain.