Predicting Amyloid Pathology in Mild Cognitive Impairment Using Radiomics Analysis of Magnetic Resonance Imaging

Background: Noninvasive identification of amyloid β (Aβ) is important in mild cognitive impairment (MCI) patients for better clinical management. This study aimed to evaluate whether radiomics features in the hippocampus in MCI improve the prediction of cerebrospinal fluid (CSF) Aβ 42 status when integrated with clinical and genetic profiles. Methods: A total of 407 MCI subjects from the Alzheimer’s Disease Neuroimaging Initiative were allocated to the training (n = 324) and test (n = 83) sets. Radiomics features (n = 214) from the bilateral hippocampi were extracted from T1-weighted images of magnetic resonance imaging (MRI). A previously defined cutoff (< 192 pg/mL) was applied for CSF Aβ 42 status. After feature selection, random forest with subsampling methods were trained to predict the CSF Aβ 42 with three models: 1) a radiomics model; 2) a clinical model based on clinical and genetic profiles including demographics, APOE ε4 genotype, and neuropsychological tests; and 3) a combined model based on radiomics and clinical profiles. The prediction performance of the classifier was validated in the test set using the area under the receiver operating characteristic curve (AUC). Results: The radiomics model identified 33 radiomics features to predict CSF Aβ 42 , which showed an AUC of 0.674 in the best performing radiomics model in the test set. The clinical model identified 6 clinical features to predict CSF Aβ 42 , which showed an AUC of 0.758 in the best performing clinical model in the test set. The combined model based on radiomics and clinical profiles identified a total of 37 features (32 from radiomics and 5 from clinical features), showing an AUC of 0.823 in the best performing combined model test set, which showed the highest performance among the three models. Conclusions: Radiomics model from MRI help predict CSF Aβ 42 status in MCI patients and potentially triage the patients for the invasive and costly Aβ test.

disease prognosis as well as for potential preventative and therapeutic treatments. (2) Therefore, biomarker-based detection of the initial amyloid β (Aβ) pathology is important for better clinical management of MCI, potentially providing the opportunity to start disease-modifying therapies before the progression of AD.
Aβ pathology can be assessed by measurement of Aβ concentration in the cerebrospinal fluid (CSF) or via molecular imaging techniques such as positron emission tomography (PET) scans using a specific radioligand for Aβ. (3) However, obtaining CSF is by lumbar puncture is invasive, and PET scans are costly, invasive due to radiation exposure, and are not always available. (4) Therefore, finding noninvasive predictive biomarkers for Aβ status could reduce the number of invasive examinations and financial burden.
Structural neuroimaging using magnetic resonance imaging (MRI) has been shown to be useful in characterizing dementia and cognitive decline due to AD pathology. (5,6) Structural changes in ADvulnerable structures such as the entorhinal cortex, hippocampus, and temporal lobe have been reported to be diagnostic indicators of cognitive impairment and even used for the prediction of amyloid pathology. (6) Compared with CSF study and PET scan, MRI has the advantages of being noninvasive and its expenditure is usually reimbursed in most countries. Therefore, if MRI can predict the Aβ pathology, it can have potential advantages over CSF study or PET scans.
Radiomics is an emerging field that extracts automated quantifications of enormous radiologic phenotype using data characterization algorithms. (7) Because radiomics models use high-throughput imaging features, they likely discover hidden information that is inaccessible with single-parameter approaches. To the best of our knowledge, there has been no previous radiomics study to predict the amyloid pathology in the MCI population. We hypothesized that radiomics features of brain MRI along with machine learning technique and in addition to clinical and genetic profiles can aid in predicting the CSF Aβ 42 status in MCI patients.

Patient population
Neuroimaging Initiative-GO (ADNI-GO) and ADNI2 database (adni.loni.usc.edu) were included in this study. The eligible patients were those who completed baseline visit and underwent MRI. Of these, we excluded those who had 1) missing demographics or neuropsychological (NP) test data (n = 68), 2) error in hippocampus masks or severe artifacts on MRI (n = 18), and 3) error in radiomics processing (n = 1). Finally, 407 patients were enrolled in this study. The enrolled patients were semi-randomly allocated to the training (n = 324) and test (n = 83) sets, with stratification for CSF Aβ status (Fig. 1

MRI acquisition
MRIs were acquired using a 3-Tesla system as per standardized protocols compatible with the ADNI.
(13) T1-weighted images were acquired using an axial three-dimensional spoiled gradient echo sequence. Axial T2 fluid-attenuated inversion recovery images were acquired. Image postprocessing and radiomics feature extraction Automated mask extraction of right and left hippocampus was performed using volBrain (https://volbrain.upv.es/), (14,15) which is a robust automatic pipeline for brain segmentation with high accuracy. (16) After denoising with an adaptive nonlocal mean filter, images were affineregistered in the Montreal Neurological Institute space using Advanced Normalization Tools software, (17) corrected for image inhomogeneities using N4, and, finally, intensity normalized. (18) Then, the hippocampus was segmented based on a multi-atlas framework combining nonlinear registration and patch-based label fusion. (19) Two experienced neuroradiologists (Y.W.P. and M.P, with 8 years and 10 years of experience, respectively) visually checked for segmentation or registration errors by overlaying each subject's native-space-transformed ROI masks onto their T1-weighted images and modified the errors.
For radiomics analysis, all images were resampled to 1-mm isovoxels across all patients. A total of 107 radiomics features (including shape; first-order features; and second-order features consisting of gray level co-occurrence matrix, gray level run-length matrix, gray-level size zone matrix, gray level dependence matrix, and neighboring gray tone difference matrix (Supplementary Table 1

Patient characteristics
The baseline characteristics and NP test results of the 407 MCI patients in the training and test sets are summarized in Table 1. In both the training and test sets, the CSF Aβ 42 + group was significantly older (p = 0.001 and p = 0.003 in the training and test set, respectively), had higher prevalence of APOE e4 carriers (p < 0.001 and p < 0.001 in the training and test set, respectively), showed higher scores in ADAS-cog (p < 0.001 and p = 0.014 in the training and test set, respectively) and lower LM I total (p < 0.001 and p = 0.008 in the training and test set, respectively) and LM II total scores (p < 0.001 and p = 0.014 in the training and test set, respectively) compared to the CSF Aβ 42 -group.   Although previous radiomics studies in the neuroradiology field have mostly been focused on neurooncology, (36)(37)(38)(39)(40) there have also been several recent studies using radiomics analysis on T1weighted images in AD. These studies have shown promising results not only in the diagnosis of AD but also in the prediction of disease progression. (41)(42)(43)(44) Radiomics features may be prone to biological validation for their correlation with disease pathology. (45) This observation is based on the hypothesis that radiomics features, especially second-order features, capture the spatial variation in signal intensity that may reflect the deposition of Aβ plaques. Further, it may extract different biological information from volume, (43,46) which is the traditional imaging biomarker of AD.
Notably, nearly all the radiomics features were retained in the combined model after the LASSO procedure in our study. This suggests that most radiomics features harbor information independent from the clinical features, which may provide added value in predicting the CSF Aβ status. However, the prediction of CSF Aβ status by radiomics features alone was not optimal, confirming the importance of clinical features. Nonetheless, our results indicate that the added value of radiomics features over clinical features is a robust method.
Our study has several limitations. First, we only included the radiomics features of the hippocampus, as previous studies showed good performance using the hippocampus mask for the classification and We obtained data from the Alzheimer's Disease Neuroimaging Initiative database (adni.loni.usc.edu).
The Alzheimer's Disease Neuroimaging Initiative was approved by the institutional review board at each site, and all participants gave their written consent.

Consent for publication
The Alzheimer's Disease Neuroimaging Initiative was approved by the institutional review board at each site, and all participants gave their written consent.

Availability of data and materials
The dataset analyzed in the current study are available in the Alzheimer's Disease Neuroimaging Initiative database.

Competing interests
The authors declare that they have no competing interests.

Funding
None.
YWP analyzed and interpreted the patient demographics and clinical data and drafted the manuscript.