Intracranial volume segmentation for neurodegenerative populations using multicentre FLAIR MRI

Abstract

Intracranial volume (ICV) segmentation, also known as brain extraction or skull-stripping, is a critical preprocessing step in analytical pipelines for studying neurodegenerative diseases in magnetic resonance imaging (MRI). While the fluid-attenuated inversion recovery (FLAIR) MRI modality has emerged as an important sequence for analyzing cerebrovascular and neurodegenerative disease, most existing automated ICV segmentation methods have been developed for T1-weighted or multi-modal inputs. Additionally, many methods have been designed using single centre data of healthy subjects and encounter difficulties using images with varying acquisition parameters and neurodegenerative pathology. In this work, we develop and evaluate 2 traditional and 8 deep learning algorithms for ICV segmentation in FLAIR MRI. Training and testing were completed on 175 ​vol (8317 images) from 2 dementia and 1 vascular disease cohort. A human phantom FLAIR MRI dataset from a repeatedly scanned, healthy individual was also utilized for reliability analysis. Images were acquired from 47 imaging centres with varying scanners and parameters. To measure and compare performance, we present a novel framework for evaluating the effectiveness of computer generated segmentations on multicentre datasets. The evaluation framework includes assessments of algorithm accuracy, generalization capabilities, robustness to pathology and spatial location, and volumetric measurement reliability - all important dimensions for establishing proof of effectiveness (a prerequisite to clinical translation). The top performing method was a multiple resolution U-Net (MultiResUNet), which achieved a mean Dice similarity coefficient greater than 98% and was robust across pathology levels and spatial locations. Our results confirm a FLAIR-based ICV analytical pipeline can alone be utilized for large-scale neurodegenerative disease research. The presented evaluation framework can be deployed by other researchers to assess the viability of tools proposed for automated analysis of diverse, clinical MRI datasets.

Keywords: Brain extraction; Deep learning; FLAIR MRI; ICV segmentation; Skull-stripping.

Fig. 1

Encoding/decoding units for U-Net, Res U-Net, and Dense U-Net. Encoding units reside between max pooling layers and decoding units reside between transposed convolutional layers.

Fig. 2

Encoding/decoding unit for MultiResUNet. Encoding units reside between max pooling layers while decoding units reside between transposed convolutional layers.

Fig. 3

Summary of the data split strategies and data preparation process prior to algorithm training and testing. All algorithms were trained, validated, and tested using the same subsets of data.

Fig. 4

Sample images showing variation of scans from different datasets and different scanner vendors.

Fig. 5

Volume histograms before and after intensity standardization. The intensity ranges used to estimate different tissue volumes via thresholding are shown by the colours on the standardized histograms. Pixels representing the brain (GM/WM), CSF, and WML are shown as turquoise, green, and blue, respectively.

Fig. 6

Sample images showing variation of scans with different CSF loads and different WML loads.

Fig. 7

Slices from SIMON volumes acquired on GE, Philips, and Siemens scanners. Despite all being acquired with the CDIP protocol (Yushkevich et al., 2006), differences in scanning hardware manifest themselves in the output volumes.

Fig. 8

Sample ICV segmentations from data split #1 models. The top rows are 3 different ADNI volumes, the middle rows are 3 different CAIN volumes, and the bottom rows are 3 different CCNA volumes. Red overlays show the ground truth (GT) annotations, green overlays show traditional method outputs, and turquoise overlays show deep learning method outputs.

Fig. 9

DSC, HD, and EF distributions across all ICV segmentation methods for data split #1.

Fig. 10

DSC, HD, and EF distributions across all ICV segmentation methods for data split #2.

Fig. 11

B-A plots between manual ground truth and algorithm predicted ICV in millilitres for data split #1 and data split #2. Mean of methods is given on the x-axis and difference between methods is given on the y-axis.

Fig. 12

DSC distributions across all ICV segmentation methods for data split #1 as a function of scanner vendor.

Fig. 13

DSC distributions across all ICV segmentation methods for data split #2 as a function of dataset.

Fig. 14

DSC distributions across select ICV segmentation methods for data split #1 as a function of WML load, CSF load, and MoCA categorization.

Fig. 15

Sample segmentations for challenging cases across select ICV segmentation methods. Red overlays show ground truth delineations, green overlays show traditional algorithm predictions, and turquoise overlays show 2D CNN predictions. The top row is a CVD case from the CAIN database, the middle row is an MCI case from the CCNA database, and the bottom row is an AD case from the ADNI database.

Fig. 16

Average regional DSC across select ICV segmentation methods. Each region represents progressive 20% increments of volume slices. Standard deviation is shown as black error bars.

Fig. 17

Error maps for select ICV segmentation methods. Averaged error maps of segmentations are shown for slices 5, 15, 25, 35, and 45 in the 55 slice registration atlas space.

Fig. 18

ICV measurements computed on SIMON as a function of method and scanner vendor.

Fig. 19

Coefficients of variation (CoV) between scanners and methods.

Fig. 20

Average intensity standardized histograms of all CAIN, ADNI, and CCNA volumes from each scanner vendor.