AbstractThe months between birth and age 2 are increasingly recognized as a period critical for neuro-development, with potentially life-long implications for cognitive functioning. However, little is known about the growth trajectories of brain structure and function across this time period. This is in large part because of insufficient approaches to analyze infant MRI scans at different months, especially brain segmentation. Addressing technical gaps in infant brain segmentation would significantly improve our capacity to efficiently measure and identify relevant infant brain structures and connectivity, and their role in long-term development. In this paper, we propose a transfer-learning approach based on convolutional neural network (CNN)-based image segmentation architecture, QuickNAT, to segment brain structures for newborns and 6-month infants separately. We pre-trained QuickNAT on auxiliary labels from a large-scale dataset, fine-tuned on manual labels, and then cross-validated the model’s performance on two separate datasets. Compared to other commonly used methods, our transfer-learning approach showed superior segmentation performance on both newborns and 6-month infants. Moreover, we demonstrated improved hippocampus segmentation performance via our approach in preterm infants.
Generalizing deep whole brain segmentation for pediatric and post-contrast MRI with augmented transfer learning
