InceptionWTMNet: A hybrid network for Alzheimer’s Disease detection using wavelet transform convolution and Mixed Local Channel Attention on finely fused multimodal images

Abstract

Multimodal fusion has emerged as a critical technique for the diagnosis of Alzheimer’s Disease (AD), with the aim of effectively extracting and utilising complementary information from diverse modalities. Current fusion methods frequently cause the precise alignment of source images and do not adequately address parallax issues. This oversight can result in artifacts during the fusion process when images are misaligned. In response to this challenge, we propose a refined registration fusion technique, termed MURF, which integrates multimodal image registration and fusion within a cohesive framework. The Vision Transformer (ViT) has inspired the application of large-kernel convolutions in the diagnosis of Alzheimer’s disease (AD) because of its ability to model long-range dependencies. This approach aims to expand the receptive field and enhance the performance of diagnostic models. Despite requiring a minimal number of floating-point operations (FLOPs), these deep operators encounter challenges associated with over-parameterisation because of high memory access costs, which ultimately compromises computational efficiency. By utilising wavelet transform convolutions (WTConv), we decompose large-kernel depth-wise convolutions into four parallel branches. One branch employs a wavelet-transform convolution with square kernels, while the other two branches incorporate orthogonal wavelet-transform kernels with an identity mapping. This innovative method, with a Mixed Local Channel Attention mechanism, has facilitated the development of the InceptionWTConvolutions network. This network maintains a receptive field comparable to that of large-kernel convolutions, while concurrently minimising over-parameterisation and enhancing computational efficiency. InceptionWTMNet classified AD, MCI, and NC using MRI and PET data from ADNI dataset with 98.69% accuracy, 98.65% recall, 98.70% F1-score, and 98.98% AUC. and provide Graphical abstract in correct format.