Early diagnosis of Alzhiemer’s disease using wavelet-pooling based deep convolutional neural network

• Published in: Sadhana (Bangalore) Vol. 48; no. 3
• Main Authors: Raju, Manu, P. Gopi, Varun, Anitha, V S, Sherawat, Abishek
• Format: Journal Article
• Language: English
• Published: New Delhi Springer India 19.08.2023
• Subjects: Engineering; multi-path CNN; mild cognitive impairment; wavelet-pooling; Alzheimer’s disease; multi-scale CNN
• ISSN: 0973-7677; 0973-7677
• DOI: 10.1007/s12046-023-02219-8

Coronal anatomic slices of structural MRI images clearly show the topographical structures of the Hippocampus and Amygdala, which are essential for early diagnosis of Alzheimer’s disease (AD). MR coronal sections are best appreciated for studying the complex topographical relationships of the amygdala and the topographical structures of the hippocampus, which helps in the early detection of disease. Early diagnosis helps prevent the disease’s progression to its final stage. It allows the patient to be aware of the severity of the disease and can take the necessary therapeutic medications to prevent its progression. A coronal view study of MR images is proposed in this paper for early diagnosis of disease using a wavelet-pooling-based multi-path and multi-scale convolutional neural network. This work aims to perform a three-way classification of 2D coronal slices of MRI images to diagnose Mild Cognitive Impairment, AD, and Normal Control in a single algorithm and learn the brain-affected regions through Gradcam visualization. wavelet-pooling is utilized to extract the texture details of the image and thus provide spatial attention to the texture features of the image, which is impossible using Max-pooling or Average-pooling. Multi-scale feature learning is incorporated using parallel multiple low-rank convolutional kernels to capture varying scales of atrophy regions. Multi-path mode compensates for the early loss of features and avoids vanishing gradient problems. The proposed model is trained and tested on the ADNI dataset comprising 900 subjects to give an accuracy of 96.5 % with ten-fold cross-validation. The multi-scale and multi-path methods significantly reduce the number of learnable parameters.