Classification of cognitively normal controls, mild cognitive impairment and Alzheimer’s disease using transfer learning approach

Abstract

Automated classification of dementia stage using imaging will be useful for clinical diagnosis and the classification accuracy will be biased for highly imbalanced samples in each class. Hence, we propose a novel approach using transfer learning-based structural significance (TLSS) for the classification of cognitively normal controls (CN), mild cognitive impairment (MCI) and Alzheimer’s disease (AD) patients based on white matter Gaussian diffusion (tensor) indices and non-Gaussian diffusion (kurtosis) indices. The structural T1-weighted magnetic resonance imaging and diffusion images were taken from ADNI dataset with 44 CN, 84 MCI and 22 AD patients. We estimate the regional Gaussian diffusion indices such as tensor fractional anisotropy (TFA) and mean diffusivity (TMD) as well as non-Gaussian diffusion indices such as kurtosis fractional anisotropy (KFA) and kurtosis mean diffusivity (KMD) in white matter regions. Further, we build transfer learning model using various balanced classifiers with structural expansion reduction (SER) and structure transfer using threshold (STT) and ensemble of majority voting of both SER and STT algorithms. We build two models by training the source model using kurtosis indices, refine the model on target tensor indices and vice versa. Transfer learning model using balanced random forest classifier was able to classify and predict all the groups with an overall accuracy about 0.79 using ensemble of SER and STT forests rather than individual algorithms (SER and STT). Our results conclude that the proposed model using kurtosis indices as source model classified and predicted with accuracies of 0.96, 0.72 and 0.7 in classifying CN vs AD, CN vs MCI and AD vs MCI respectively. To conclude, the proposed approach has improved the classification accuracy and its potential applicability for imbalanced data sample datasets.