Controls-based denoising of voxel-based morphometry and multimodal imaging data improves prediction of conversion to Alzheimer’s disease

Ziel/Aim We recently suggested controls-based denoising (CODE), a novel approach to improve medical image analysis by identifying patterns of physiological variance from healthy controls (HC) using principal component analysis and removing them from patient data. When estimating the extent of a pathological effect with a pattern expression score (PES), CODE is likely to reduce the error. We previously showed that CODE improves prediction of conversion to Alzheimer’s disease (AD) with FDG-PET. Here, we investigated the extent to which CODE improves conversion prediction based on voxel-based morphometry (VBM) and multimodal image analysis.

Methodik/Methods T1-weighted MRI and FDG-PET were obtained from 133 healthy controls (HC), 87 AD patients and 206 patients with mild cognitive impairment (MCI; 119 AD-converter, conversion within 4 years and 87 non-converter, minimal follow-up 4 years) from ADNI database and spatially normalized using DARTEL. From T1-weighted MRI we calculated VBM images. The PES of an AD-pattern was calculated for each image modality separately. Classification performance was evaluated for PES-VBM, PES-FDG and for a combination (logistic regression) using area under curve (AUC), Matthew correlation coefficient (MCC), sensitivity (SENS) and specificity (SPEC).

Ergebnisse/Results Based on a linear model, we estimated that CODE increases the signal-to-noise ratio of PES-VBM by a factor of 2.16. Classification performance of PES-VBM increased from AUC 0.74 to 0.81 (p = 0.007), MCC 0.36 to 0.53, SENS 0.68 to 0.75 and SPEC 0.69 to 0.78. AUC of PES-FDG increased from 0.80 to 0.85 (p = 0.046). In logistic regression, PES-FDG was the dominant predictor (p < 0.001), but PES-VBM also contributed significantly (p = 0.017), yielding a classification performance of AUC 0.86 and MCC 0.70 (as opposed to MCC 0.63 with PET only).

Schlussfolgerungen/Conclusions Using CODE to improve prediction of AD-conversion from VBM data, we achieved a higher performance than most analyses reported in the literature. In multimodal analysis, FDG-PET remains the dominant predictor, with MRI also contributing significantly.