For many years, a special case of neuroimaging methods known as resting-state functional magnetic resonance imaging (rs-fMRI) has been recognized as the benchmark for diagnosing regional and time-varying fluctuations in brain metabolism by measuring alterations in cerebral blood flow (He et al., 2024; Arpanahi et al., 2024). This approach indirectly gauges blood oxygen level-dependent (BOLD) signals within the human brain, subsequently utilized to visualize and quantify brain activities. Given that rs-fMRI is capable of identifying neuropathological alterations associated with various diseases, its application in the clinical diagnosis of Alzheimer's disease (AD) and mild cognitive impairment (MCI) is on the rise. Consequently, it serves as a biomarker for the pathology and advancement of AD and helps to distinguish Alzheimer's from other neurodegenerative disorders (Shanmugavadivel et al., 2023). Nonetheless, within the realm of digital pathology, brain specimens can be examined on a computer monitor as whole-slide images, produced by digitizing rs-fMRI scans with high-resolution imaging devices. Owing to progress in digital pathology and the growing availability of extensive public datasets, neuropathology has emerged as a significant field for applying data-driven models, including deep neural networks (DNN). Deep learning techniques have been utilized for various purposes, including object detection (Shahzad and Ali, 2024; Chamakuri and Janapana, 2024), image categorization (Alorf and Khan, 2022; Loddo et al., 2022), and the segmentation of cerebral regions (Arpanahi et al., 2024; Hu et al., 2022). Moreover, deep networks have been employed to tackle more intricate challenges, such as predicting genetic mutations and conducting survival assessments. Furthermore, pre-trained convolutional neural networks (CNN) are utilized as feature extractors in content-based image retrieval (CBIR) (Asadi Amiri et al., 2022).

The growing integration of artificial intelligence (AI) within healthcare, particularly in pathology, underscores the necessity for medical practitioners to comprehend the functioning of AI in clinical assessments. Nevertheless, a study involving British medical students revealed they perceive themselves as insufficiently equipped to employ medical AI tools (Sit et al., 2020). Consequently, it is essential for healthcare practitioners and institutions to enhance their ability to comprehend and effectively utilize the potential of these advancing technologies. A deeper grasp of AI can foster greater trust and acceptance of these innovations, enabling healthcare professionals to use these tools more proficiently in medical realms (Wiljer, 2020; Thompson et al., 2018; Aktolun, 2019).

A key skill required to grasp the functioning of AI is the capability to identify and articulate the decision-making processes of AI systems (Long and Magerko, 2020). A significant concern with many present deep learning models is that their decision-making mechanisms are often hidden from users, commonly known as the "black box" issue. Additionally, classification models are typically developed using extensive publicly accessible datasets (Purushotham et al., 2018). However, in neuropathology, these datasets are unlikely to comprehensively cover all brain regions, disease categories, and various sample preparations (for instance, differing imaging planes like axial, coronal, and sagittal, as well as variations in image resolution, etc.). For example, the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset comprises rs-fMRI data from six categories: cognitively normal (CN), significant memory concern (SMC), early mild cognitive impairment (EMCI), mild cognitive impairment (MCI), late mild cognitive impairment (LMCI), and patients with AD. Therefore, when creating an AI algorithm to diagnose AD through rs-fMRI images, it is crucial to comprehend precisely what the algorithm has assimilated.

These challenges prompted us to investigate the decision-making process of VGG-16, a pre-trained DNN designed for classifying stages of Alzheimer's through feature extraction (Tammina, 2019). Consequently, ongoing research focuses on developing an innovative approach for the early diagnosis of AD, utilizing deep feature extraction from rs-fMRI images of individuals with and without AD. More specifically, this research aims to address the following inquiry:

Can the deep features elicited by VGG-16 from rs-fMRI images serve as a practical foundation for diagnosing AD?

In the context of neuropathology images, initiatives have been undertaken to enhance the interpretability of DNNs by investigating deep features. In general, it is desirable to reduce the number of deep features to a selection of informative ones. Faust et al. demonstrated that the intricate individual characteristics of a DNN trained on images of brain tumors correspond to identifiable histomorphological patterns. This aspect of deep features was examined by analyzing deep feature activation maps (Faust et al., 2020; Faust et al., 2019). Schomberg et al. identified significant deep features by randomizing the values of deep features and assessing how this randomization impacted the classification accuracy of their model (Schaumberg et al., 2020). In the current research, the classification performance of each deep feature is examined separately from the other deep features. It aims to assess the interpretability of deep features produced by a pre-trained DNN model when utilizing rs-fMRI data from CN, SMC, EMCI, MCI, LMCI, and AD groups. This research may assist professionals in the field who lack expertise in AI to assess models intended for integration into their processes.

The subsequent sections of this paper are structured as follows. Section 2 provides a concise overview of the database employed and the quantitative analyses applied in this research. Section 3 details the experiments and results derived from the database. Section 4 discusses the experimental findings and evaluates the effectiveness of the proposed approach. Finally, Section 5 wraps up the paper with the conclusion and future works.