Source-linked AI summary
Multidimensional classification of hippocampal shape features discriminates Alzheimer's disease and mild cognitive impairment from normal aging
Emilie Gerardin, Gaël Chételat, Marie Chupin, Rémi Cuingnet, Béatrice Desgranges, Ho-Sung Kim, Marc Niethammer, Bruno Dubois, Stéphane Lehéricy, Line Garnero, Francis Eustache, Olivier Colliot
TL;DR
The paper asks whether hippocampal shape can improve automatic discrimination of AD and MCI from healthy aging, beyond limitations of global volumetry and prior whole-brain approaches. It represents automatically segmented hippocampi with spherical harmonics, selects features using bagging, and classifies them with SVMs. The method achieved 94% accuracy for AD versus controls and 83% for MCI versus controls, with higher accuracy than reported hippocampal volumetry.
Problem
Early AD diagnosis, particularly for MCI, is challenging; volumetry captures global hippocampal change, while prior classifiers lacked detailed hippocampal morphological analysis.
Method
The method automatically segments hippocampi, represents their surfaces with SPHARM coefficients, selects relevant features through bagging, and classifies patients with SVMs.
Results
94% accuracy was achieved for AD versus controls and 83% for MCI versus controls, exceeding reported hippocampal volumetry accuracy.
Takeaways & Limitations
Multidimensional hippocampal shape classification may assist Alzheimer’s disease diagnosis within the evaluated study scope.
Takeaways & Limitations
The results require confirmation in larger participant groups, and leave-one-out accuracies may be optimistic because parameter and feature selection occurred outside the cross-validation loop.
Abstract
from arXiv · showhide
We describe a new method to automatically discriminate between patients with Alzheimer's disease (AD) or mild cognitive impairment (MCI) and elderly controls, based on multidimensional classification of hippocampal shape features. This approach uses spherical harmonics (SPHARM) coefficients to model the shape of the hippocampi, which are segmented from magnetic resonance images (MRI) using a fully automatic method that we previously developed. SPHARM coefficients are used as features in a classification procedure based on support vector machines (SVM). The most relevant features for classification are selected using a bagging strategy. We evaluate the accuracy of our method in a group of 23 patients with AD (10 males, 13 females, age $\pm$ standard-deviation (SD) = 73 $\pm$ 6 years, mini-mental score (MMS) = 24.4 $\pm$ 2.8), 23 patients with amnestic MCI (10 males, 13 females, age $\pm$ SD = 74 $\pm$ 8 years, MMS = 27.3 $\pm$ 1.4) and 25 elderly healthy controls (13 males, 12 females, age $\pm$ SD = 64 $\pm$ 8 years), using leave-one-out cross-validation. For AD vs controls, we obtain a correct classification rate of 94%, a sensitivity of 96%, and a specificity of 92%. For MCI vs controls, we obtain a classification rate of 83%, a sensitivity of 83%, and a specificity of 84%. This accuracy is superior to that of hippocampal volumetry and is comparable to recently published SVM-based whole-brain classification methods, which relied on a different strategy. This new method may become a useful tool to assist in the diagnosis of Alzheimer's disease.
Introduction
Early Alzheimer’s disease can be difficult to diagnose, especially at the MCI stage, motivating MRI-based analysis of hippocampal changes. This paper addresses limitations of volumetry and prior whole-brain classifiers by using multidimensional hippocampal shape features with SVMs.
- Motivation: Early AD diagnosis is challenging, particularly for amnestic MCI patients who may represent a transition to early Alzheimer’s disease.MCI involves cognitive deficits while preserving independent living.
- Motivation: High-resolution 3D MRI can reveal subtle anatomical changes, and the hippocampus is affected by pathological changes at early AD stages.Prior MRI studies assessed hippocampal atrophy using manual segmentation and volumetry.
- Research gap: Volumetry assesses global hippocampal changes, whereas shape analysis can reveal local atrophy and may be more sensitive, especially in MCI.Prior shape studies reported prominent atrophy in regions corresponding to CA1 and the subiculum.
- Research gap: Prior multidimensional classifiers used anatomical features from distributed whole-brain regions but did not provide detailed hippocampal morphological analysis.The hippocampus may provide relevant classification information because it is affected early in the pathological process.
- Contribution: The proposed method automatically discriminates AD and MCI patients from healthy controls using hippocampal shape features represented by spherical harmonics and classified with SVMs.Spherical harmonics provide a parametric 3D boundary description analogous to Fourier series.
Materials and methods
The study combines automatic hippocampal segmentation, spherical-harmonic surface representation, feature selection, and SVM classification. It evaluates these methods on AD, MCI, and control participants using MRI and leave-one-out cross-validation.
- Participants and imaging: The sample included 23 AD patients, 23 amnestic MCI patients, and 25 elderly healthy controls.Participants underwent T1-weighted MRI scans, with demographic and clinical characteristics reported for the patient groups.
- Hippocampal segmentation: Hippocampi were segmented automatically using a previously developed method based on competitive region-growing between the hippocampus and amygdala.The method incorporates probabilistic-atlas priors and relative-position information.
- SPHARM representation: Hippocampal surfaces were represented with SPHARM-PDM, which maps spherical harmonics onto surfaces with spherical topology.The workflow converts segmentations to meshes, computes a spherical parameterization, truncates the series at degree L=20, and normalizes for rotation and translation.
- Feature construction: SPHARM coefficients served as classification features, while SPHARM-PDM landmarks visualized the localization of shape differences.Each subject was represented by two sets of three-dimensional coefficients, one for each hippocampus.
- Feature construction: A degree-20 decomposition produced a 2646-feature vector by concatenating three coordinates across both hippocampi.The feature count is 2×3×(20+1)^2=2646.
- Feature selection: Feature selection used Student’s t-tests with bagging, retained the most discriminative features, and then centered and normalized them with z-scores.The bagging procedure repeatedly computed statistics on resampled subject sets and kept features with consistently significant differences.
- Classification: SVM classification used patient and control labels, with linear and radial-basis-function kernels evaluated using cost parameter C and kernel-width parameter γ.The RBF kernel was selected as the primary nonlinear classification option and compared with a linear kernel.
- Validation: Classification accuracy, error, sensitivity, and specificity were computed with leave-one-out cross-validation, including feature selection within each training loop.A grid search estimated the SVM parameters C and γ, and accuracy was evaluated across different numbers of selected features.
Results
Hippocampal SPHARM classification distinguished AD and MCI from controls, with strongest performance for AD and consistently better MCI performance than comparator approaches.
- Classification performance: 94% correct classification, 96% sensitivity, and 92% specificity were achieved for AD versus controls using 16–22 features.The optimal feature count corresponded to a p-value of 0.002.
- Classification performance: 83% correct classification, 83% sensitivity, and 84% specificity were achieved for MCI versus controls using 2–3 features.The optimal feature count also corresponded to a p-value of 0.002.
- Parameter influence: Accuracy was stable across maximal SPHARM decomposition degrees for MCI and reached 94% for AD at L=15 and L=20.AD accuracy was 92% at L=5 and L=10; MCI accuracy was 83% for all tested degrees.
- Comparison with voxel-based SVM: On the ADNI group, the method reached 88% accuracy for AD versus controls and 80% for MCI versus controls.For AD, Klöppel’s voxel-based method reached 90% accuracy; for MCI, it reached 71%.
- Comparison with SPHARM-PDM: SPHARM coefficients matched SPHARM-PDM for AD classification but achieved higher MCI accuracy: 83% versus 75% with univariate selection and 69% with PCA.The SPHARM-PDM PCA result used five features for MCI classification.
- Comparison with prior SPHARM methods: Shen et al.’s SPHARM-PDM method reached 92% and 96% accuracy for left and right hippocampi in AD versus controls, but 67% and 73% for MCI versus controls.The MCI accuracies were lower than those obtained with the present approach.
- Shape differences: The most significant AD-related deformations occurred in the medial hippocampal head and a region approximately corresponding to CA1, while CA2 and CA3 were relatively spared.MCI showed similar but less extended deformation patterns.
Discussion
The paper presents SPHARM-based hippocampal shape classification with SVMs for individual discrimination of AD, MCI, and controls. It reports high accuracy, compares shape-based classification with volumetry and alternative classifiers, and notes validation and sample-size limitations.
- Method: SPHARM coefficients provide multiscale hippocampal shape features for multidimensional SVM classification, unlike univariate measures that ignore relationships between coefficients.The coefficients represent different levels of shape detail and were selected for classification features.
- Classification performance: 94% accuracy was achieved for AD versus controls and 83% for MCI versus controls, exceeding reported hippocampal volumetry performance.The paper attributes the higher discriminative power to incorporating shape deformations rather than only global volume.
- Feature selection: AD showed more discriminative features than MCI at the same statistical threshold, while adding less relevant features reduced accuracy, particularly for MCI.Optimal feature counts were 16–22 for AD and 2–3 for MCI at approximately p=0.002.
- Comparative evaluation: The approach outperformed SPHARM-PDM and FLD alternatives for MCI-versus-control classification, suggesting coefficients were more efficient than landmarks for this task.The comparison was conducted against SVM with SPHARM-PDM and the method of Shen et al.
- Shape localization: The hippocampal deformation pattern approximately corresponded to CA1 in AD and was similar but less extended in MCI, although MRI resolution prevented direct subfield mapping.CA2 and CA3 were relatively spared in the reported localization analysis.
- Limitations: The results require confirmation in larger, age-matched samples, and the study could not assess conversion from MCI to AD because too few converters were available.Controls were significantly younger than the AD and MCI groups, and longitudinal confirmation was recommended.
Tables
Table 1 compares the paper’s approach with Klöppel et al.’s method on the ADNI group.
- Table 1 presents a comparison between the paper’s approach and Klöppel et al.’s 2008 method.
- The comparison is conducted on the ADNI group.
- The table is intended to compare classification approaches rather than hippocampal morphology alone.
Figures
The figures show classification performance, parameter selection, feature-driven shape variation, and regional hippocampal differences for AD and MCI versus controls.
- AD versus controls: 94% accuracy is achieved for AD versus controls with an RBF-kernel SVM using 16–22 selected features.The parameter-performance figure reports the highest accuracy in this feature range.
- MCI versus controls: 83% accuracy is achieved for MCI versus controls with an RBF-kernel SVM using 2–3 selected features.The highest accuracy occurs with substantially fewer features than in the AD comparison.
- Parameter selection: Grid searches evaluate accuracy, error, sensitivity, and specificity across SVM cost C and kernel-width γ values.The AD results are shown for a 19-feature classifier, while the MCI results are shown for a 3-feature classifier.
- Shape variation: The two most discriminative SPHARM coefficients are illustrated by adding or subtracting 1SD and 2SD from the mean hippocampal shape.This visualizes shape modes associated with AD versus controls classification.
- Regional differences: AD-related deformations significantly affect the medial head and lateral body of the hippocampus, approximately corresponding to CA1.Significant atrophy is shown at p<0.05 with multiple-comparison-corrected P-values.
- Regional differences: MCI and control hippocampi show an atrophy pattern similar to AD versus controls but covering a smaller spatial extent.The figure uses corrected significance maps to display regional group differences.