Abstract
The accomplishment of recognizing Alzheimer’s disease AD in its beginning stages especially identifying mild cognitive impairment MCI will cause the onset of the disease to occur later. This is very tough to perform due to the finesses associated with making early pathological changes visually undifferentiated. MRI imaging provides helpful information about the structural problems associated with AD, yet many current computer-aided diagnostic systems do not have fine enough representation of features or precise spatial location information. Therefore In response to these difficulties the authors propose EDA-Net Enhanced Deep Attention Network a newly developed Deep convolutional Neural Network to detect multiple stages of AD from MR images. In particular we are proposing an innovative component of EDA-Net called the Deep Attention DA Block which utilizes our newly developed Efficient Coordinate-Spatial Attention Mechanism ECSAM that incorporates Efficient Channel Attention ECA and Coordinate Attention CA and provides a mechanism for sequentially connecting the various channels of the DA Block with respect to collecting the inter-channel information as well as modeling the long range spatial dependencies within the network to focus dynamically on the areas of greatest discrimination in the areas of the MRI images. The DA Block uses SiLU activation and Group Normalization GN to provide both additional nonlinear representational ability and stability when training the network with small-batches of training data. EDA-Net has been validated through five multiple folds of cross-validation on a publicly available database containing 6400 MR images obtained from four consecutive clinical stages. Based on experimental results, EDA-Net is able to achieve a mean classification accuracy of 98.30%, with all other important metrics having an accuracy of over 98.38%, and a Kappa coefficient of 97.73%. This essentially makes EDA-Net superior in performance to traditional models, such as ResNet50 and DenseNet121. The results of the ablation studies have shown a synergy between ECA and CA components, and the confusion matrix has confirmed that this network has a very high level of class discrimination. Overall, EDA-Net has a unique architecture combined with a hybrid mechanism of attention to allow it to balance between high accuracy, efficient use of resources, and optimum performance. For all of these reasons, EDA-Net is a very robust deep learning solution to screen for the early stages of Alzheimer's disease. Future research will continue validating the generalizability of the model using more diverse clinical datasets and will be investigating ways to integrate different types of information sources to enable clinical translation.