Semantic Neuroanatomical Segmentation of Brain MRI Scans Using Pretrained SynthSeg Neural Network

Deep 3D convolutional neural networks (CNNs) with trained algorithms have proven the possibility to distinctively partition brain regions to reduce manual division and expert labeling of neuroanatomical structures. Although many machine learning techniques implement this strategy, most can be time-consuming and quite difficult to interpret as contrast modeling with varying resolutions can be obscure upon differentiating across regions. SynthSeg, a 3D U-Net algorithm, contains the ability to segment scans in datasets of multiple modalities and a variety of target domains without the need to retrain as it is trained with preconstructed data derived from a general 3D U-Net model for volumetric segmentation, thus facilitating the analysis of large quantities of bioimaging data in short periods of time. This is made possible due to the randomization technique where the contrast and resolution of the preconstructed training data are selected randomly to increase deep learning performance and accuracy. Due to its ability to learn from fluctuating contrast and resolution types via the randomization technique, its segmentation robustness is maintained across datasets. SynthSeg can learn from pretrained designated labels for characterization of diverse populations to improve detection and diagnoses of a vast array of diseases. Using SynthSeg, a subset of the Child and Adolescent NeuroDevelopment Initiative (CANDI) dataset containing brain MRI volume for a single patient was used with its corresponding ground truth image to analyze and exhibit highly accurate predictive modeling within each scan. The average Dice index across all labels for the MRI volume is 0.80793, which indicates an almost perfect segmentation accuracy.