Automatic U-Net based Segmentation Pipeline for Neonatal and Child Brain MRI

Pouria Mashouri, Fateme Sadat Haghpanah, Sam Osia, Min Sheng, Mireille Guillot, Jarred Garfinkle, Dianna McAllister, Michael Brudno, Ting Guo

Keywords: Segmentation, Computer Vision, MRI, Radiology, Neonatal, Child

TL;DR: A robust U-Net model for the automatic brain segmentation of child MRIs, ranging from preterm to 8-year children

Abstract:

Background: Robust segmentation on magnetic resonance images (MRI) is key to assessing brain growth, which is critical to the health and development of children. Machine learning has enabled breakthroughs in automatic brain MRI segmentation, albeit mainly for adults. Accurate delineation of infant brain structures is more challenging, given the significant growth and dramatic anatomical variability of the brain during the first years of life.

Objective: We present a U-Net based automatic segmentation pipeline to robustly segment brain structures on both neonatal and child MR images with enhanced robustness of data input.

Methods: A total of 300 T1-weighted images were included: 100 early-in-life, 100 term equivalent age (TEA), and 100 8-year (median age at MRI: 32 weeks, 40.3 weeks, and 8.2 years respectively). Each image was accompanied with manually segmented annotations that cover cerebrum, cerebellum, and brainstem. We broke up each 3D image volume along the sagittal axis to 256 2D slices of 256x256 matrices. The U-Net model took as input a single slice, and returned a multi-class prediction denoting the mask of each label. The resultant 256 slices were then re-constructed back to a 3D scan. Our model was trained on a 60/20/20 training/validation/testing split across the age range for 200 epochs.

Results: When compared to the manually segmented labels, our pipeline achieved an average DICE coefficient of 97.78% (95.40%-98.47%), 94.85% (89.58%-97.91%), and 94.84% (88.39%-97.12%) for cerebrum, cerebellum, and brainstem respectively, and outperformed InfantFS and FreeSurfer, applied to the same data. More importantly, our model had a robust performance across different age groups, demonstrating its applicability to a much wider age range than other existing state-of-the-art methods, as well as robust performance on an independently acquired external cohort.

Conclusions: Our U-Net based pipeline allows accurate segmentation of brain structures across the first years of life. It is a flexible, robust, and automated tool to assess total and regional brain volumes and growth for studying healthy children and children with medical conditions.

Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.

Paper Type: methodological development

Primary Subject Area: Segmentation

Secondary Subject Area: Image Acquisition and Reconstruction

Confidentiality And Author Instructions: I read the call for papers and author instructions. I acknowledge that exceeding the page limit and/or altering the latex template can result in desk rejection.