Abstract: Cognitive impairment arising from ischemic stroke, Alzheimer’s disease, and Parkinson’s disease presents distinct structural and network-level alterations. Brain magnetic resonance imaging offers a non-invasive and high-resolution approach to assess these changes, while deep learning provides powerful tools for automated analysis. Given that accurate lesion delineation, precise localization of abnormal regions, and reliable disease classification are fundamental to clinical decision-making. This review aims to explore the application of deep learning techniques to brain magnetic resonance imaging analysis of cognitive impairments caused by these disorders, with a focus on three core tasks: lesion segmentation, object detection, and image classification. Recent widely accepted findings indicate that ischemic stroke studies have achieved state-of-the-art lesion segmentation performance, with optimized U-shaped convolutional network (U-Net) and hybrid convolutional neural network-transformer models reaching Dice scores up to 0.911 in delineating focal damage. Alzheimer’s disease research has advanced classification and staging accuracy by more than 10% compared with unimodal baselines through three-dimensional convolutional neural network, Transformers, and multimodal fusion, enabling more precise detection of diffuse cortical atrophy. Parkinson’s disease imaging, despite lacking overt structural lesions, has leveraged ResNet and Vision Transformer backbones to identify subtle and spatially distributed abnormalities, improving early-stage differentiation. Persistent challenges include the scarcity of large, high-quality annotated datasets, substantial inter-site variability, high annotation costs, and limited interpretability, hindering clinical integration. Addressing these barriers will require advances in federated learning to mitigate data scarcity while preserving privacy, domain adaptation techniques to reduce inter-site variability, automated annotation, and low-resource training strategies to lower labeling costs, and explainable artificial intelligence to improve interpretability, thereby ensuring model robustness, privacy, and transparency. This review highlights emerging methods, innovative technologies, and novel paradigms that are redefining brain imaging analysis in cognitive impairment. Mechanistically, deep learning improves cognitive impairment analysis by integrating hierarchical and multiscale spatial features, modeling long-range functional connectivity disruptions, and fusing structural with functional imaging to better represent network-level pathology. In conclusion, aligning network architectures with disease-specific imaging characteristics and task requirements can greatly enhance the accuracy, robustness, and generalizability of magnetic resonance imaging analyses for cognitive impairment. Future work should focus on multimodal fusion, structure-function coupling, cross-disease evaluations, and embedding artificial intelligence tools into clinical workflows to support early detection, individualized treatment planning, and large-scale clinical adoption.

Key words: Alzheimer’s disease, cognitive dysfunction, deep learning, image classification, ischemic stroke, lesion segmentation, magnetic resonance imaging, neuroimaging, object detection, Parkinson’s disease