Executive Summary : | The automated segmentation of pediatric brain magnetic resonance images, particularly those of the neonatal and infant brains, is a clinically significant yet challenging area due to its low signal-to-noise ratio, age-specific changes, high inter-subject variability, and limited availability of atlases and segmentation tools. This project aims to perform tissue-wise segmentation and volumetric estimation of these brains using a transfer learning method using a convolutional neural network. The problem with pediatric brain segmentation is the lack of thoroughly annotated atlas images and ground truths. To address this issue, the proposed method uses transfer learning based on adult brain MRI segmentation, using the large availability of adult brain atlases and segmentation labels for training. This pretrained model will then be applied to the neonatal and infant brains, incorporating changes in the pediatric brain caused by ongoing brain development and insufficient myelination based on the baby's age. The study will also consider preterm-born neonates and infants with congenital abnormalities as special cases to assess their brain growth and development level. The obtained 2D and 3D segmentation results will be compared qualitatively and quantitatively with ground truth images using metrics such as dice ratio, Jaccard index, Hausdorff distance, sensitivity, and specificity. A tissue-wise comparison of results will also be made with recently published papers in the state-of-the-art literature. |