A comparison of accurate automatic hippocampal segmentation methods

Abstract

The hippocampus is one of the first brain structures affected by Alzheimer's disease (AD). While many automatic methods for hippocampal segmentation exist, few studies have compared them on the same data. In this study, we compare four fully automated hippocampal segmentation methods in terms of their conformity with manual segmentation and their ability to be used as an AD biomarker in clinical settings. We also apply error correction to the four automatic segmentation methods, and complete a comprehensive validation to investigate differences between the methods. The effect size and classification performance is measured for AD versus normal control (NC) groups and for stable mild cognitive impairment (sMCI) versus progressive mild cognitive impairment (pMCI) groups. Our study shows that the nonlinear patch-based segmentation method with error correction is the most accurate automatic segmentation method and yields the most conformity with manual segmentation (κ=0.894). The largest effect size between AD versus NC and sMCI versus pMCI is produced by FreeSurfer with error correction. We further show that, using only hippocampal volume, age, and sex as features, the area under the receiver operating characteristic curve reaches up to 0.8813 for AD versus NC and 0.6451 for sMCI versus pMCI. However, the automatic segmentation methods are not significantly different in their performance.

Accurate and robust segmentation of neuroanatomy in T1‐weighted MRI by combining spatial priors with deep convolutional neural networks

“…Twenty subjects were selected from each of the following clinical subgroups: normal controls, mild cognitive impairment, and Alzheimer's disease. Since this dataset was previously used to compare several state‐of‐the‐art algorithms (Zandifar et al, ), for our experiments, we use the same data (e.g., previously preprocessed as described in Zandifar et al) to enable meaningful comparisons with the results reported in the aforementioned work. Pre‐processing consisted of patch‐based (PB) denoising (Coupé et al, ), N3 nonuniformity correction (Sled, Zijdenbos, & Evans, ), linear intensity normalization to the range [0,100], and affine registration to the MNI‐ICBM152 template (Fonov et al, ) with 1 × 1 × 1 mm 3 resolution.…”

“…We further compared several variants of our CNN‐based method with several other popular and/or state‐of‐the‐art segmentation methods on the same dataset using the segmentations previously produced in the work of Zandifar et al (), which includes results for four different methods, both before and after applying EC (Wang et al, ), a machine learning based wrapper which attempts to correct systematic errors made by the initial host segmentation method. The methods included are FreeSurfer (Fischl, ), ANIMAL (Collins & Pruessner, ) (a multiatlas technique combining nonlinear registration with majority‐vote label fusion), traditional PB segmentation (Coupé et al, ) (a multiatlas technique combining linear registration with PB label fusion), and an augmented approach combining PB segmentation with nonlinear registration.…”

“…Note that the results from the 2.5D CNN method (Kushibar et al, 2018) are not included in the boxplots because only summary statistics were available. Of our CNN-based methods, the performance of the baseline network CNN-B was poorest overall T A B L E 5 Comparison of four of our CNN-based segmentation methods with previously reported results (Zandifar et al, 2017)…”

“…Some such methods have attempted to learn complex mappings between image features and labels using traditional machine-learning based classifiers (e.g., support vector machines (Boser, Guyon, & Vapnik, 1992) and random forests (Breiman, 2001)) combined with handcrafted feature sets (Morra et al, 2010;Zikic et al, 2012), while others have found success transferring labels using a combination of linear or nonlinear image registration with local and/or nonlocal label fusion (so-called "multiatlas segmentation" methods (Coupé et al, 2011, Heckemann, Hajnal, Aljabar, Rueckert, & Hammers, 2006, Iglesias & Sabuncu, 2015). Indeed, many state-of-the-art results (e.g., hippocampus segmentation (Zandifar, Fonov, Coupé, Pruessner, & Collins, 2017) and brain extraction (Novosad & Collins, 2018) exploit a complementary combination of both multiatlas segmentation and machine-learning methods (e.g., error correction (EC) (Wang et al, 2011)).…”

“…The left hippocampus is overlaid in green, and the right hippocampus in blue. Errors are overlaid in red in columns four and six Compared to the best method fromZandifar et al (2017), which combines PB segmentation with nonlinear registration and EC (PBS + NLR + EC), our best performing method (CNN-SP-D + DA) yielded an improvement of 2.1% in terms of mean Dice and a decrease in mean MHD of 0.17 mm (over both left and right hippocampi), both of which were highly statistically significant (p ≤ 10 −9 ). CNN-SP-D + DA was also considerably more robust than the methods examined in the work of Zandifar et al, producing fewer outliers with low overlap (Figure 5).…”