Advances in brain imaging have generated unprecedented volumes of high-dimensional data, yet extracting meaningful information from complex, noisy, and incomplete brain imaging data remains a significant challenge. Diffusion models (DMs) have introduced a paradigm shift in this field, surpassing traditional generative approaches. This review systematically examines the theoretical foundations of diffusion models, and their practical applications in eight brain imaging computing tasks: registration, super-resolution, cross-modal reconstruction and synthesis, segmentation, classification, brain network analysis, brain-computer interface (BCI) signals augmentation, and BCI decoding. Additionally, we emphasize obstacles that hinder deployment in practice, including computational scalability and sampling inefficiency, limited generalization under domain shift sensitivity, as well as multimodal integration and alignment constraints, while outlining potential future directions that emphasize the convergence of diffusion models with large-scale foundation models, which hold the potential to advance scalable, reliable, and clinically embedded brain imaging solutions. Throughout this review, we aim to establish a roadmap of progress and translational hurdles to guide emerging research and accelerate collaboration spanning DMs, clinical brain imaging, and engineering disciplines.

In Rapid Serial Visual Presentation (RSVP), the vast majority of stimuli are not consciously perceived, but the salient ones breakthrough into awareness and can be reported. In addition, these breakthrough events are observable with EEG, since they generate a P3 or other distinguishing components. The Fringe-P3 method is based upon these characteristics. Concealed knowledge studies have successfully employed this Fringe-P3 method using own-name and own email-address, with the method being shown to be less vulnerable to counter-measures than other approaches. It has also been shown that famous faces presented in RSVP differentially break into awareness and generate a distinct evoked response component. In this paper, we further enhance the applicability of the Fringe-P3 method by demonstrating the effectiveness of the method on personally-familiar faces. While salient, such stimuli do not have the exquisite salience of famous faces, being a better match to the level of salience that might be found in forensic applications. Our findings suggest that the Fringe-P3 method could be used to detect intrinsic salience of familiar faces, even when there was no task associated with these faces. We investigated the sensitivity of the ERP-based RSVP paradigm to infer recognition of familiar faces, and performed statistical inference in the Time and Frequency domains, to differentiate between known and unknown faces, at group and participant levels.

Accurate detection of Mild Cognitive Impairment (MCI) is critical for timely intervention and for slowing progression to Alzheimer's disease. Electroencephalography (EEG) offers a non-invasive and cost-effective measure of brain activity; however, its complex, non-linear dynamics limit conventional analysis. We propose a CNN-Res-SE-BiLSTM-BiGRU framework for the automated detection of MCI directly from raw EEG. Convolutional and residual blocks capture local temporal structure, bidirectional recurrent layers model long-range dependencies, and Squeeze-and-Excitation (SE) modules provide channel-wise attention. Predicted probabilities are calibrated using temperature scaling, and operating thresholds are selected on the validation set using Youden's J statistic. The model is evaluated using five-fold cross-validation under both subject-dependent and strict subject-independent protocols on a primary resting-state dataset, with additional subject-independent validation on an odor EEG dataset. Under subject-independent evaluation on the odor dataset, the proposed model achieved an accuracy of 0.956 ± 0.051, with ROC-AUC of 0.971 ± 0.051 and PR-AUC of 0.934 ± 0.132. UMAP-based visualization and explainable AI analyses (SHAP and LIME) provide interpretable insight into the learned spatiotemporal patterns and sample-specific decisions. These results demonstrate robust, interpretable EEG-based MCI detection with potential clinical utility.

This article presents a robust and efficient framework for brain tumor segmentation based on deep learning. We introduce a novel three-dimensional (3D) mass-preserving geometric transformation (MPGT) that employs a homotopy method to transform irregular brain magnetic resonance (MR) images into standardized solid cubes. This transformation preserves local mass ratios while maintaining global structural integrity, providing a structured input for deep learning models. Furthermore, we propose a modified two-phase segmentation strategy to minimize inference time and a postprocessing technique to enhance lesion-wise performance. Extensive validation on the Brain Tumor Segmentation (BraTS) Challenge 2023 dataset demonstrates that our method, when integrated with nnU-Net, achieves competitive Dice scores of 0.9282 (Whole Tumor), 0.8812 (Tumor Core), and 0.8527 (Enhanced Tumor). These results are superior to or comparable with top-ranking competition entries.

The rapid evolution of machine learning (ML) methods has yielded promising results in human brain neuroscience. However, the reproducibility of ML applications in neuroimaging remains limited, challenging the generalizability of inferences to broader populations. In addition to the inherent variability of the brain activity (both in healthy and pathological states), poor reproducibility is further enhanced by inconsistencies in data preprocessing techniques and methods for calculating functional connectivity (FC), which are used as parameters for brain state classification. To systematically assess the impact of abovementioned factors on ML applications to fMRI data, we benchmarked a comprehensive set of FC analysis pipelines for the classification task between fMRI data recorded in two fundamentally different states: eyes open and eyes closed. In contrast to studies involving heterogeneous clinical populations or using complex cognitive tasks, our controlled experimental design - based on two independent datasets of healthy participants collected in different laboratories - minimizes variability related to a task design or pathological brain states. Classification accuracy and reproducibility were compared for 256 distinct FC analysis pipelines, covering common preprocessing approaches, brain parcellation schemes, and connectivity metrics. Notably, we employed two ways of validation: a direct cross-site validation strategy - when a model was trained on one site and tested on another, and few-shot domain adaptation - when a few samples of testing site were added to the train set. Despite the substantial variability in pipeline configurations, we observed consistently high classification accuracy (~ 90%), confirming that FC-based models can robustly discriminate between well-defined brain states (eye conditions) across different acquisition sites. Best results both in terms of classification accuracy and stability were observed using Pearson correlation and tangent space parametrization as FC, Brainnetome as atlas, and confound regression strategies based on the CompCor method. These findings highlight the resilience of rs-fMRI FC-derived characteristics to methodological variation and support their utility in the discovery of biomarkers, particularly in settings that involve stable and reproducible brain states.

The classification of functional brain networks plays an important role in the diagnosis of neurodegenerative diseases, brain decoding and other fields. Functional brain networks can effectively reflect the functional connection relationships between brain regions or neurons and accurately represent brain activities. Therefore, a large number of problems related to the classification of functional brain networks have been studied. However, the traditional functional brain network merely measures the static correlation between brain regions or neurons in a simple way, and does not reflect the causal transmission effect between brain regions. This directionality is crucial for the regulatory relationship between brain regions. Furthermore, since the brain is constantly in a state of dynamic change, the dynamics of functional connectivity also plays a very important role in the classification of functional brain networks. Therefore, we propose a classification framework named Dynamic Directed Propagation Networks (DDPN) for functional brain networks considering the dynamic directed propagation mechanism. This method effectively captures the dynamics and directionality of the dynamic directed brain network and further improves the classification accuracy of the functional brain network. To verify the effectiveness of the proposed method, we conduct experiments on real datasets. The experiments show that the proposed method improved by 3.1-4.1% compared with state-of-the art methods in two datasets.

The noise of Magnetic Resonance Imaging (MRI) poses challenges for Deep Learning (DL) when tumor boundaries are obscured, tumor location and appearance are complex due to overlap between tumor and non-tumor cells, and modality identification is difficult because tumor features vanish in the later layers of the DL. Effective feature extraction from given MRI is a possible solution to overcome this challenge. Therefore, we develop BrainFusionNet that combines Convolutional Neural Networks (CNNs), Vision Transformers (ViT), and Gated Recurrent Units (GRUs) to extract spatial, contextual, and sequential features from MRI images for improved brain tumor classification. Furthermore, explainable AI such as SHAP, LIME, and Grad-CAM are integrated to visualise and highlight image regions that contribute to BrainFusionNet's decision-making process. The proposed BrainFusionNet model is evaluated on two publicly available MRI datasets. K-fold validation suggests 98% accuracy on both datasets. The model was compared with the six state-of-the-art (SOTA) CNNs and transfer learning. Among the SOTA CNNs, DenseNet121 and VGG16 achieved the highest accuracy of 96%. The novelty of BrainFusionNet is that the hybrid model effectively extracts local and global features from MRI images, even in small-scale tumor regions and small tumor sizes. The model has a balanced sequential CNN architecture to capture low-level and deeper-layer features; a customized ViT that captures local features, stabilizes gradient flow, and reduces the risk of vanishing gradients during MRI image training. The CNN and ViT outputs are fed into a GRU for final classification. Furthermore, we analyze pixel intensities to determine whether MRI image quality affects image classification. Our findings are very novel in image interpretation, as we found that the distribution of pixel intensities in MRI images affects DL performance.

MELD Graph is a state-of-the-art artificial intelligence (AI) model for automated detection of focal cortical dysplasia (FCD), but its performance remains limited, highlighting the need to investigate which aspects of the pipeline affect its accuracy. A retrospective failure-mode analysis of the MELD Graph pipeline was performed in 242 subjects, with model predictions and FreeSurfer segmentations reviewed to classify errors as segmentation-associated or algorithm-related. FCD imaging features salient to humans were quantified, with statistical associations examined for both MELD Graph detection and focal FreeSurfer segmentation failure. MELD Graph demonstrated overall performance similar to previously published non-harmonized results, achieving a sensitivity of 69%, specificity of 44%, and positive predictive value (PPV) of 75%. Focal FreeSurfer segmentation failures were associated with 21% of false negative patients, 25% of false positive clusters in patients, and 16% of false positive clusters in controls. Following manual cortical segmentation correction and rerunning of MELD Graph, 67% of the segmentation-associated missed lesions were detected, and segmentation-associated false positive clusters were reduced or eliminated in 75% of controls with such clusters. Higher conspicuity on T1-weighted images was associated with MELD Graph detection, whereas greater conspicuity on T2-FLAIR images relative to T1 was associated with detection failure. Non-bottom-of-sulcus lesion location, higher human conspicuity measures, and low T1 image quality were positively associated with focal FreeSurfer segmentation failures. FreeSurfer segmentation failures are a significant potential source of error in the MELD Graph pipeline. FCD imaging features salient to humans and image quality were also associated with variability in algorithm performance. Robust cortical segmentation and stronger integration of T2-FLAIR imaging features may be beneficial for automated FCD detection tools. Not applicable. This study is a retrospective analysis of previously acquired open-source imaging datasets and does not constitute a clinical trial.