A Deep Learning Pipeline for Automatic Skull Stripping and Brain Segmentation

Abstract

A deep learning pipeline (DLP) with a triple network framework was developed to perform skull stripping and segment the brain into gray matter, white matter and cerebrospinal fluid (CSF) using T1w Magnetic Resonance (MR) images. Three separate 3D Dense-Unets were designed to decompose the complex skull stripping and brain segmentation problems into individual binary segmentation problems to segment a particular label using a 32x32x32 patch based approach. These included a skull stripping network to obtain the brain mask (BM), GW-net to segment gray matter (GM) and white matter (WM), and CSF-net to segment cerebrospinal fluid (CSF). The networks consisted of seven dense blocks with each block containing four layers. Every layer was connected to every other layer in that dense block. Each layer consisted of four sublayers namely, BatchNormalization, 3D Convolution, ReLu and dropout. As a part of the iTAKL study [1], 785 T1w MR datasets including 288 high school (14-18 years) and 497 youth (9-13 years) datasets were used. On the evaluation dataset of 50 held-out subjects, dice scores of (a) 0.980, 0.92, 0.94 and 0.845 for BM, GM, WM and CSF respectively on down sampled data, and (b) 0.983, 0.9103, 0.9277 and 0.83 for BM, GM, WM and CSF respectively on the full resolution data were achieved. The pipeline was then tested on datasets from the AADHS study and 5 other studies from the Human Connectome project (HCP) [2, 3] with comparable performance.