FULLY AUTOMATIC WHOLE BRAIN LONGITUDINAL SEGMENTATION, VALIDATION AND OPERATIONALIZATION

Abstract

Whole-brain atrophy is a well established biomarker for clinical trials in AD and other neurodegenerative diseases. Widely used methods of whole-brain atrophy estimation, such as the boundary shift integral (BSI) (Freeborough; IEEE, 1997), require manual intervention. As such, these methods are susceptible to intra/inter-segmentor variability and require trained resource for quality control (QC). Here we present a fully automated pipeline for estimation of whole-brain atrophy with subsequent automatic QC testing, minimising sources of potential variability and need for visual inspection. We employed a subset (n=372) of the ADNI1 cohort with 1.5T T1 images and KMN-BSI values available (Fox Lab, ADNI) at Baseline, 12 month and a 24 month follow-up as validation data set. Baseline images were brain extracted with PinCram (Heckemann;PLOS,2015) and N4 Bias field corrected (Tustison;IEEE,2010). Longitudinal images were registered to the baseline image, bias corrected and intensity normalised to the baseline intensity profile. Segmentation was performed on baseline images using a whole-brain multi-atlas segmentation approach (WB LEAP; Wolz, NeuroImage, 2010; Ledig, Proc ISBI,2012). From this, we obtain subject probabilistic maps of cerebral tissue types used to initialise a probabilistic temporospatial segmentation using expectation maximisation based optimisation (Ledig, IEEE,2014). To ensure data quality is maintained, a visual inspection of the input baseline tissue segmentations is required only. QC of image registrations and longitudinal segmentations is performed automatically through tolerance testing on the spatial cross-correlation between registered images, annualised atrophy and normalised volumetric change. Resultant atrophy estimations were compared to KMN-BSI values for validation. The proposed automatic QC reported a 90% pass rate for longitudinal segmentations; upon inspection, all automatic QC pass grades were found to be correct. Segmentations failing automatic QC were visually inspected, 8 subjects failed due to poor registrations. Good agreement was observed between atrophy estimated with KNM-BSI and the proposed method; significant correlation was observed in atrophy estimations at month 12 (r2=0.62, p<<0.01) and month 24 (r2=0.73,p<<0.01). Reported effect and sample sizes were not significantly different between methods for any clinical group or follow-up. We have proposed a fully automatic method for longitudinal assessment of whole-brain atrophy with semi-automatic QC, reducing potential segmentor/rater variance and reporting atrophy consistent with the BSI method.