Fully automated segmentation of the locus coeruleus: application to assess the effects of age, sex and education

BackgroundLocus coeruleus (LC) imaging using neuromelanin-sensitive (NM) MRI sequences is a promising biomarker for detecting early Alzheimers Disease (AD) and Parkinsons Disease (PD). Although semi or fully automatic approaches have been developed to estimate LC integrity by measuring its intensity, these techniques most often rely on a single template built in a standardized space and/or depend on a number of voxels to be accounted that is defined a priori. Thus, these algorithms make it impossible to perform direct volumetric analyses and do not properly account for inter-individual anatomical variability. To fill this gap, our aim was to develop a new multi-atlas fully automated segmentation method using the Automatic Segmentation of Hippocampal Subfields (ASHS) software. A second aim was to investigate the effects of age, sex and education on LC intensity and volume. MethodWe used cross-sectional data from 101 cognitively unimpaired older adults (mean age: 73.8{+/-}6.6 years; mean education: 13.2{+/-}3.0 years; 58 women, 43 men) from the Age-Well randomized controlled trial for whom high-resolution NM-MRI (T1-w with magnetization transfer; 0.3x0.3x0.75mm3) and standard T1-w MRI (1x1x1mm3) were available. The LC were manually segmented in 30 randomly selected participants on NM-MRI, and the standard T1-w MRI, NM-MRI and bilateral segmentations were fed into the ASHS training pipeline to generate a new atlas (ASHS-LC). ASHS-LC was applied to the 71 remaining subjects to segment the LC and we assessed the effects of age, sex and education on both i) LC intensity (normalized by the intensity of the pons) and ii) LC volume (normalized by the total intracranial volume). ResultFive-fold cross-validation experiments revealed high accuracy of the automatic segmentation relative to manual segmentation (Dice coefficient 0,83{+/-}0,04). ICCs demonstrated excellent reliability for intensity (ICC = 0.99), whereas reliability for volume was lower (ICC = 0.43). LC intensity was significantly higher in women than in men (Cohens d = 0.94, p < 0.001) while no associations with age ({beta} = -0.0002, p = 0.98) or education ({beta} = 0.11, p = 0.37) were found. In contrast, LC volume was not different between men and women (Cohens d = 0.10, p= 0.55) but was significantly negatively associated with age ({beta} = -0.25, p = 0.04) and education ({beta} = -0.27, p = 0.04). Lastly, LC intensity and volume were not statistically correlated, neither using manual segmentation (atlas set: {beta} = 0.05, p = 0.78) nor ASHS segmentation (atlas set: {beta} = 0.11, p = 0.57; analysis set: {beta} = -0.07, p = 0.55). ConclusionOverall, this new method allows to automatically and accurately segment the LC and offers the opportunity to measure its integrity both in terms of intensity and volume. This is of importance since these two metrics may offer complementary insights into LC integrity, as evidenced by the differing influences of demographic factors on LC intensity and volume.