Spatially Attentive Scale Invariant Feature Modelling for Alzheimer’s Disease Detection

The prevalence of Alzheimer's Disease (AD) patients has significantly increased recently, and early diagnosis is crucial. AI-aided solutions for diagnosing AD have been rising for decades. The majority of these solutions show the application of deep learning methods, both with and without transfer learning. However, a few issues have been noted that require attention in order to detect AD: a. there is a lack of scale-invariant feature modelling, and b. there is a lack of spatial attention paid to scale-invariant features. The suggested model addresses these issues in two ways: first, an efficient multiscale feature modelling module (EMFMM) is designed to exploit scale invariant features through multilayers each having multicolumn of dilated convolution layers. At each layer the multicolumn features are concatenated and forwarded to its subsequent layer. Secondly, the spatial convolution block attention (SCBA) is plugged in different layers of concatenated features to provide spatial attention over the multiscale features. The spatially attentive feature maps from several SCBAs are further concatenated and given to a multilayer neural network to employ disease forecasting. Two open-source datasets, OASIS and AD-Dataset, are used to demonstrate extensive experiment. The suggested model performs better than current best practices, with accuracy values on OASIS and AD-Dataset of 99.75% and 96.20%, respectively.