Partially adaptive optimization driven spatial focused CNN with Gompertz non-linearity for interpretable Alzheimer’s disease diagnosis
Muhammad Waqar, Zeshan Aslam Khan, Mirza Hashim Ali Baig, Chung-Chian Hsu, Ihsan Ul Haq, Saadia Khan, Chuan-Yu Chang, Naveed Ishtiaq Chaudhary, Muhammad Asif Zahoor Raja
Recently, deep learning has revolutionized various scientific disciplines. Strategies based on deep learning have consistently surpassed traditional methods, proving extraordinary efficiency in the healthcare environment. Alzheimer’s disease is one of the major global health threats due to its rapid increase and late diagnosis. Originating from complicated neuroanatomical variations, Alzheimer’s disease is caused by the brain’s accumulation of amyloid and tau proteins. Various deep learning methods have been proposed in the literature, yet these models demand significant time and computational resources due to their multi-layered architectures. Therefore, a solution providing a good balance between accuracy and computational efficiency is a dire need of time. The proposed research contributes in the following directions, (1) the transformer-inspired spatial attention mechanism-driven lightweight convolutional neural network (CNN) model is proposed to produce an accurate and efficient solution for Alzheimer’s disease diagnosis, (2) furthermore, a non-linear activation function named Gompertz Linear Unit is exploited in the proposed network for capturing complex relationship in the given data and to overcome the issues of dead neuron and vanishing gradient faced in existing activation operations, (3) to address the challenges of over adaptiveness in existing techniques, a partially adaptive variant of Adam optimizer is utilized in this study. Moreover, the incorporation of partial adaptivity provides the flexibility to fine-tune the model as needed which improves the performance in terms of accuracy and speedy convergence, (4) explainable artificial intelligence (XAI) is exploited to produce an interpretable diagnostic insight which will be valuable for healthcare professionals to make the informed clinicals decisions. The proposed model achieves a substantial test accuracy of 99% on OASIS database. Moreover, the proposed approach outclasses the existing benchmark models in terms of both accuracy and computational cost. Seeing the emergence of the proposed solution in accurately and efficiently diagnosing Alzheimer’s disease, it is depicted that the proposed solution has potential to serve as a smart healthcare system for detection and classification of Alzheimer’s disease at premature stages.