Deep learning model for infant brain tissue segmentation using 3D patch-wise U-Net with weighted-average reconstruction

Abstract

Low tissue contrast and motion errors make it difficult to accurately segment gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) from baby brain MRI. For automated segmentation of baby brain tissues, this study proposes a 3D patch-wise U-Net architecture with a weighted-averaging reconstruction approach. When the technique was tested on the iSeg-2017 dataset, it produced Dice scores of 0.92 (GM), 0.80 (WM), and 0.74 (CSF) with outstanding boundary accuracy—HD95 of 1.0 mm, 1.4 mm, and 2.1 mm, and ASD of 0.25 mm, 0.36 mm, and 0.48 mm, respectively. The approach creates spatially coherent segmentations appropriate for quantitative neurodevelopmental study by striking a balance between computational economy and anatomical precision. These results show the model’s promise as a trustworthy tool for newborn brain imaging research and therapeutic applications.