TL;DR
This study leverages radiomics analysis of multiparametric MRI to identify potential biomarkers for preclinical Alzheimer's Disease (AD). The researchers developed classification models that can distinguish individuals in the preclinical AD stage with high accuracy, and identified specific radiomic features that correlate with amyloid-beta deposition and predict future cognitive decline.
What Problem Does This Paper Address?
Diagnosing Alzheimer's Disease (AD) in its earliest, preclinical stages is crucial for effective intervention. However, current diagnostic methods like amyloid PET scans are costly and invasive, and there's a lack of convenient, non-invasive biomarkers for early detection. This gap makes it challenging to identify individuals at high risk before significant cognitive impairment occurs. The paper aims to address this by exploring whether radiomics features extracted from multiparametric MRI (MPMRI) can serve as these much-needed preclinical AD biomarkers.
Methods
The study employed a multi-cohort approach to train, validate, and test their radiomics models. Key methodological steps included:
- Multiparametric MRI (MPMRI) Acquisition and Preprocessing: Structural MRI (sMRI), functional MRI (fMRI), and diffusion tensor imaging (DTI) data were acquired and preprocessed using standard pipelines (e.g., DPARSF, PANDA, SPM12) to ensure image quality and consistency.
- Radiomics Feature Extraction: A vast number of features (over 30,000) were extracted from each MRI modality, including texture and wavelet features for sMRI and fMRI, and diffusion parameters for DTI.
- Feature Selection: A novel, robust feature selection strategy was implemented. This involved a combination of statistical tests (t-tests, autocorrelation) and three independent algorithms (Fisher score, Lasso, mRMR) applied with five-fold cross-validation and 100 repetitions to identify stable, high-frequency features that were less susceptible to perturbations.
- Classification Models: Support Vector Machine (SVM) and Random Forest (RF) models were trained and validated using the selected radiomics features to classify individuals into preclinical AD and control groups.
- Longitudinal Analysis: Stable high-frequency features were further analyzed in a separate cohort to assess their changes over time with cognitive decline and their impact on conversion time using paired t-tests and survival analyses.
- Correlation Analysis: The identified radiomics features were correlated with amyloid-beta deposition (measured by SUVR from PET scans) to understand their biological relevance.
Datasets
The study utilized data from three distinct cohorts:
- Cohort 1 (SILCODE): 183 cognitively healthy participants with amyloid-PET imaging. This cohort was split into an 80% training set and a 20% validation set, with participants categorized as preclinical AD (amyloid-positive) or controls (amyloid-negative).
- Cohort 2 (SILCODE): 51 participants selected retrospectively from the SILCODE project. This cohort was divided into "converters" (those who later developed cognitive decline) and "nonconverters" (those who remained cognitively stable) and served as a separate test set.
- Cohort 3: 37 participants, comprising 24 converters from Cohort 2 and 13 converters from the Alzheimer's Disease Neuroimaging Initiative (ADNI). This cohort was used for independent longitudinal validation.
Key Insights and Ideas
- Radiomics as a Powerful Tool for Preclinical AD: The study strongly suggests that radiomics features from MPMRI can effectively identify individuals in the preclinical stage of AD, a stage where traditional clinical and volumetric measures are often insufficient.
- Identification of Stable, High-Frequency Features: The novel feature selection methodology successfully identified a small set of "stable high-frequency" radiomics features that were consistent across different selection algorithms and cross-validation runs. This robustness is key for biomarker development.
- Correlation with Amyloid-Beta Deposition: The identified stable features showed significant correlations with amyloid-beta deposition, a hallmark of early AD pathology. This provides a crucial link between the imaging features and the underlying biological process.
- Predictive Power for Cognitive Decline: These radiomics features not only distinguished preclinical AD but also demonstrated the ability to predict future cognitive decline, with some features showing significant predictive power in survival analyses.
- Structural MRI Dominance: Interestingly, the most stable and predictive radiomics features were derived solely from structural MRI (sMRI), highlighting its potential in early AD detection, even when fMRI and DTI were also analyzed.
Key Results
- Classification Accuracy: The SVM and RF models achieved excellent classification performance, with average accuracies ranging from 81.9-95.9% in the validation set and 83.2-89.1% in the test set. This significantly outperforms models based on pure clinical data.
- Predictive Biomarkers: Three stable high-frequency features were identified:
- Large zone high-gray-level emphasis (LZHGE) of the right posterior cingulate gyrus (sMRI).
- Variance feature of the left superior parietal gyrus (sMRI).
- Coarseness feature of the left posterior cingulate gyrus (sMRI).
- Correlation with Aβ: Feature 6489 (coarseness) was positively correlated with mean cortical SUVR (r = 0.433, p < 0.0001), while features 11517 (LZHGE) and 27442 (variance) were inversely correlated (r = -0.416 and r = -0.348, respectively, p < 0.0001).
- Longitudinal Impact: The variance feature (27442) of the left superior parietal gyrus showed decreased levels with cognitive decline and was found to affect conversion time (p < 0.05).
Limitations and Future Work
- Small Sample Size: The authors acknowledge that the limited sample size, particularly in cohort 3, restricts statistical power and generalizability. They suggest multicenter collaboration to increase sample size.
- Definition of Preclinical AD: The study used a definition of "unstable preclinical AD" based on a 36-month window for unstable MCI, which could be refined.
- Imaging Protocol Variability: Differences in imaging protocols across sites could influence model performance.
- Limited Scope of fMRI/DTI Features: The study notes that while they explored fMRI and DTI, the stable features identified were solely from sMRI, suggesting further investigation into fMRI and DTI radiomics might be warranted.
- Lack of Tau Information: The latest NIA-AA 2018 definition for preclinical AD includes tau deposition, which was not assessed in this study.
- Future Directions: The authors propose that radiomics analysis of MPMRI holds promise as a new evaluation method for Aβ deposition and predicting future cognitive decline, crucial for ultra-early secondary prevention trials. They also highlight the need for further validation and exploration of novel feature extraction paradigms.
Summary updated