Dual-Model Radiomic Biomarkers Predict Development of Mild Cognitive Impairment Progression to Alzheimer's Disease

Abstract

Predicting progression of mild cognitive impairment (MCI) to Alzheimer's disease (AD) is clinically important. In this study, we propose a dual-model radiomic analysis with multivariate Cox proportional hazards regression models to investigate promising risk factors associated with MCI conversion to AD. T1 structural magnetic resonance imaging (MRI) and ¹⁸F-Fluorodeoxyglucose (FDG) positron emission tomography (PET) data, from the AD Neuroimaging Initiative database, were collected from 131 patients with MCI who converted to AD within 3 years and 132 patients with MCI without conversion within 3 years. These subjects were randomly partition into 70% training dataset and 30% test dataset with multiple times. We fused MRI and PET images by wavelet method. In a subset of subjects, a group comparison was performed using a two-sample t-test to determine regions of interest (ROIs) associated with MCI conversion. 172 radiomic features from ROIs for each individual were established using a published radiomics tool. Finally, L1-penalized Cox model was constructed and Harrell's C index (C-index) was used to evaluate prediction accuracy of the model. To evaluate the efficacy of our proposed method, we used a same analysis framework to evaluate MRI and PET data separately. We constructed prognostic Cox models with: clinical data, MRI images, PET images, fused MRI/PET images, and clinical variables and fused MRI/PET images in combination. The experimental results showed that captured ROIs significantly associated with conversion to AD, such as gray matter atrophy in the bilateral hippocampus and hypometabolism in the temporoparietal cortex. Imaging model (MRI/PET/fused) provided significant enhancement in prediction of conversion compared to clinical models, especially the fused-modality Cox model. Moreover, the combination of fused-modality imaging and clinical variables resulted in the greatest accuracy of prediction. The average C-index for the clinical/MRI/PET/fused/combined model in the test dataset was 0.69, 0.73, 0.73 and 0.75, and 0.78, respectively. These results suggested that a combination of radiomic analysis and Cox model analyses could be used successfully in survival analysis and may be powerful tools for personalized precision medicine patients with potential to undergo conversion from MCI to AD.

Keywords: Alzheimer’s disease; Cox model; image fusion; mild cognitive impairment; radiomics.

FIGURE 1

The overall framework of the experimental process in this study.

FIGURE 2

Results of image fusion and MCI conversion-related ROIs. Example of fusion (C) of a T1-weighted structural MRI scan (A) and an FDG-PET scan (B); results were generated using xjView9.6 Slice Viewer. (D) Projection map of the voxel-wise two-sample t-test of GM images conducted to assess differences between MCI-c and MCI-nc. Relative reduced GM volume in MCI-c relative to MCI-nc was represented by ‘cool’ colors; relative increased GM volume in MCI-c relative to MCI-nc was represented by ‘hot’ colors (p < 0.01 FDR corrected, extent threshold ≥50 voxels). (E) Projection map of metabolic difference in MCI-c relative to MCI-nc using PET images. Relative reduced glucose metabolism was represented by ‘cool’ colors; Relative hypermetabolism was depicted by ‘hot’ colors (p < 0.01 FDR corrected, extent threshold ≥50 voxels). (F) Projection map of volume difference in MCI-c relative to MCI-nc using fused MRI/PET images. Relative reduced volume was represented by ‘cool’ colors; relative increased volume was represented by ‘hot’ colors (p < 0.01 FDR corrected, extent threshold ≥50 voxels).

FIGURE 3

The medians and interquartile ranges of differences in performance assessment indicators between relevant pairs. (A) Comparison of C-index differences between relevant pairs in the training dataset. (B) Comparison of C-index differences between relevant pairs in the test dataset. (C) Comparison of differences of relative risk stability between relevant pairs in the test dataset.