Alzheimer's disease (AD) is an irreversible, progressive, neurodegenerative illness that affects primarily older individuals. AD is characterized by a higher impairment in memory or cognitive skills than healthy adults of the same age (Neugroschl and Wang, 2011). The accumulation of abnormal amyloid-β and hyperphosphorylated tau proteins is a pathological feature of AD. Amyloid-β deposition is thought to be the cause of pathological tau formation and subsequent neurodegeneration (Jack et al., 2010; Hardy and Selkoe, 2002). The neurodegeneration leads to molecular neuropathological abnormalities in distinct neuronal brain networks, which results brain network dysfunction (Drzezga, 2018). Understanding the impairments and structural modifications in white matter connectivity, that lead to the formation of these networks, can help to identify the brain network dysfunction and structural biomarkers of AD. The presence and destruction of structural connections in structural brain networks could explain simultaneous molecular, metabolic, and functional alterations (Mito et al., 2018). According to the theory of the AD brain network, the disease attacks susceptible areas in the brain and spreads across intrinsic networks, presumably via specific white matter pathways (Raj et al., 2012; Zhou et al., 2012).

Understanding brain structural connectivity is key to elucidate how neurons and neural networks process information. The disruption of structural and functional connectivity in Alzheimer's disease has always been linked to the structural brain network destruction. Functional MR imaging has been extensively used in the study of functional connectivity analysis in AD, revealing distinct patterns of functional connectivity impairments (Mondragón et al., 2021; Tang et al., 2021; Sorg et al., 2009). Changes in white matter pathways are likely to mediate this functional network dysfunction; however, due to the difficulties of modelling complex white matter structures, impairments in specific structural connectivity between the cortical regions have not been thoroughly examined in the literature using advanced techniques. In this study, we applied deep neural network classification decisions’ explanation to comprehensively investigate the discriminative structural connectivity changes in subjects with AD.

Diffusion MR imaging is now the only approach available to analyze structural changes associated with fiber pathways in vivo and non-invasively. Several diffusion MRI based investigations have shown structural alterations in white matter that have occurred over the development of AD in the last decade, the findings of which have been presented in a number of comprehensive reviews (Chua et al., 2008; Mak et al., 2017). Despite hopeful findings of white matter changes in AD using diffusion models, quantitative analysis of FA and MD has major flaws, making their findings unreliable and anatomically difficult to interpret. Using whole-brain tract-based spatial statistical (TBSS) analysis (Acosta-Cabronero et al., 2010; Bosch, 2012) or directly analyzing individual fiber bundle methods (Daianu et al., 2016; Bendlin et al., 2010) has its own set of challenges, such as automatically segmenting white matter into known fiber bundles, quantifying properties and similarity of a bundle, analyzing a specific bundle for a subject group, and the occurrence of redundant and non-existent fibers or false positives (Campbell et al., 2005) in whole brain tractography.

The structural brain network (Sporns et al., 2005) provides a more appropriate solution, which represents the complete map of the white matter connectivity in the brain. The network not only includes edges as a list of linked regions, but also provides the weight of each connection (Hagmann et al., 2008). Structural brain networks not only have the ability to shed light on the insights of structural connectivity (Srivishagan et al., 2020) but also uncover new information about the principles that govern how distinct functional subunits are organized and interact with one another (Passingham, 2013) and pathological brain conditions (Griffa et al., 2013). Building on our preliminary studies (Subaramya et al., 2021), we used structural brain networks of AD and healthy elderly people to classify and explain the discriminative differences in white matter. Rather than analyzing individual fibers or bundles, we focused on the white matter pathways between pairs of cortical regions, which are naturally provided by the structural brain network.

Deep Neural Networks (DNNs) have produced state-of-the-art outcomes in a variety of medical imaging applications, including the identification of Alzheimer's disease using neuroimaging data (for a review, see Ebrahimighahnavieh et al., 2020; Vieira et al., 2017; Jo et al., 2019). However, the DNNs’ decisions are frequently seen as non-transparent (Castelvecchi, 2016), making it challenging to use these algorithms in clinical practice. Recently, several researchers (Zhou et al., 2016, Selvaraju et al., 2017) proposed techniques to visually explain the DNNs’ decisions in various tasks. However, only a very few recent studies have explained DNNs’ decisions in neuroimaging-based AD classification with different visualization approaches (Böhle et al., 2019; Rieke et al., 2018; Yang et al., 2018). These studies are focused on visually identifying the most influential brain regions in diagnosing Alzheimer's disease based on the decisions of 3D MR image classification. To the best of our knowledge, no research has been done to visually identify the discriminative white matter connectivity changes between cortical regions using structural brain networks.

In this study, we focus on exploring the most influential white matter connectivity changes in AD by visually explaining the convolutional neural network (CNN) decisions in structural brain network based classification. By feeding the features of the structural brain network, a CNN architecture was proposed to distinguish AD from healthy normal subjects. Then the Gradient-weighted Class Activation Mapping (Grad-CAM) technique (Selvaraju et al., 2017) was utilized to visually interpret the classifier's decision. The study investigates inside the black box of classification for AD and explains the CNN decisions regarding which changes in the structural connections will have the most impact on the classification outcome. We show that the overall approach succeeded in illustrating the discriminative pattern of white matter connectivity changes in AD. The discriminative white matter pathways have piqued the interest of researchers, and they are now being used as biomarkers for Alzheimer's disease. The outcome of this study contributes to AD diagnosis and also provides clinicians more faith in automated AD diagnostic systems.