Autism spectrum disorder is characterized by substantial heterogeneity across biological burden, adaptive reserve, developmental timing, therapeutic engagement, and intervention responsiveness. This heterogeneity increasingly motivates the development of multimodal and artificial intelligence-supported approaches in autism research. However, many existing approaches emphasize diagnostic classification, symptom prediction, or generic outcome modeling, while fewer frameworks explicitly preserve clinically interpretable constructs linking biological burden, developmental state, therapeutic accessibility, and responsiveness to intervention. This Hypothesis and Theory article proposes FIAP®-Digital as a conceptual, hypothesis-generating, and human-supervised multimodal AI architecture for future translational stratification research in autism. FIAP®-Digital is designed to integrate multimodal inputs-including biological, physiological, developmental, contextual, therapeutic-process, and longitudinal indicators-into structured estimates of FIAP® constructs such as the Biological Burden Index, energetic capacity, adaptive neurodevelopmental window accessibility, therapeutic engagement accessibility, neuroplastic accessibility, and higher-order translational profiles. The proposed architecture is construct-preserving rather than purely predictive. It emphasizes multimodal data organization, construct-level representation, temporal updating, uncertainty visibility, explainability, clinician-in-the-loop interpretation, and Responsible AI governance. The manuscript outlines candidate computational components, including multimodal fusion, temporal modeling, missing-data handling, uncertainty estimation, explainability mechanisms, and clinician-readable outputs. It also proposes operational computational representations of FIAP® constructs and a stepwise human-supervised profiling workflow. Importantly, FIAP®-Digital is not presented as a validated diagnostic tool, treatment recommendation system, autonomous AI model, or clinically proven platform. The constructs described remain hypothetical and require empirical validation. The article therefore provides a neuroinformatics architecture and validation roadmap, including feasibility testing, construct validity, internal and external validation, longitudinal robustness, clinician usability, fairness assessment, and uncertainty calibration. Its primary contribution is to define a responsible computational pathway for testing whether FIAP® constructs can support future precision-stratified autism research.
IntroductionAutomatic seizure detection from long-term scalp electroencephalography (EEG) can reduce the burden of manual EEG review. However, multi-derivation EEG is commonly processed as a collection of input channels without explicitly modeling the structured relations defined by the bipolar montage.MethodsIn this study, we propose MGF-Net, a montage-guided fusion network for seizure detection from long-term scalp EEG. MGF-Net represents 18 fixed bipolar derivations as nodes in a relation-labeled graph. The graph encodes same-chain adjacency, left-right correspondence, and shared-electrode relations derived from the bipolar montage. After epoch-wise robust normalization and wavelet-based signal reconstruction are applied independently to each epoch, a shared one-dimensional convolutional neural network extracts temporal features from each derivation. A relation-aware graph Transformer then performs feature fusion across derivations using relation-specific key and value projections. Attention-based graph pooling generates an epoch representation for ictal probability estimation, and post-processing converts epoch-wise probabilities into seizure-event detections.ResultsExperiments on the CHB-MIT dataset show that MGF-Net achieves a segment-level accuracy of 98.38%, sensitivity of 93.18%, and specificity of 98.41%. At the event level, the proposed method detected all 65 annotated test events, achieving a sensitivity of 100.00%, a false detection rate of 0.82/h, and a latency of 2.03 s.DiscussionThese results demonstrate MGF-Net's effectiveness for seizure detection from long-term scalp EEG.
To address the challenges of inconsistent inter-modal contributions, high reliance on labeled data, and limited cross-modal synergy in multimodal emotion recognition, this paper proposes a dual-stream Siamese Self-supervised Emotion Recognition method (DSSER) based on EEG and fNIRS signals. During the pretraining phase, DSSER uses a dual-stream Siamese architecture to extract spatiotemporal features from electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) signals. It also aligns their representation spaces through a co-training strategy. This approach eliminates the need for negative samples in conventional contrastive learning. In the fine-tuning stage, a Dynamic Gated Cross-Attention Fusion module is introduced. It assigns modality weights and integrates complementary cross-modal features. Experimental results on a ternary emotion dataset involving 16 participants show that DSSER achieves an accuracy of 98.03% under an intra-subject emotion-recognition protocol. This result reflects performance under the adopted intra-subject setting. It should not be interpreted as subject-independent generalization to unseen participants. The results support the effectiveness of DSSER in improving cross-modal interaction.
Decision making (DM) requires coordination of elementary information processes subserved by a distributed network of brain areas. Computational models help to understand these processes, but most of the existing models focus on simulating only one of the many parallel operations. An existing spiking neural network (SNN) model attempts to simulate DM holistically, however it does not take advantage of the significant role of inhibition at the neural level as a possible mechanism underlying DM. To address this limitation, we propose to examine the impact of neural inhibition on decision strategy selection in value-based DM using the mentioned model. In this study we outline the methodology and perform successful in-silico validation of the inhibition hypothesis with the SNN model of DM. To perform the simulation, we use a well-studied multi-attribute choice task and we validate simulation results against human behavioral data. The inhibition model achieved approximately 17% lower mean prediction error than the no-inhibition model (0.55 vs. 0.67) when evaluated on held-out, compensatory-condition data not used for fitting (Wilcoxon signed-rank test, p = 0.009, r = 0.55), with no significant difference observed in the condition used for fitting. These findings indicate that the advantage conferred by inhibition is not attributable to model complexity alone, and support neural inhibition as a plausible, biologically grounded mechanism for adaptive, context-sensitive decision strategy selection.
IntroductionAutomatically localizing the Anterior Commissure (AC) and Posterior Commissure (PC) is foundational for CT-based algorithmic disease screening, yet robust computational methods for this on CT remain lacking. We developed a registration-guided 3D-UNet framework for CT-based AC-PC localization, demonstrating its utility in computing ventriculomegaly features for Normal Pressure Hydrocephalus (NPH) detection.MethodsFramework development and evaluation were on an internal cohort of scans from patients with NPH, Alzheimer’s disease, post-traumatic volume loss, and headache (Veterans Affairs [VA]-Cohort, n = 427). External validation was on separate datasets (VA-ExtCohort, University of California, Santa Barbara [UCSB]-ExtCohort). AC-PC reference standard definition, model development, and evaluation were on 1 mm3–resampled scans.ResultsOn 1-mm3 resampled scans, test-set AC-PC mean radial errors (MREs) were 1.64/1.49 mm on the VA-Cohort, 2.42/1.79 mm on the VA-ExtCohort (n = 40), and 2.31/1.93 mm on the UCSB-ExtCohort (n = 43). Notably, the upper limits of the 95% confidence intervals (CIs) for localization errors across all cohorts were well below 3.2 mm; we empirically determined this to be a clinically relevant threshold beyond which the discriminative power of AC-PC-referenced ventriculomegaly features degrades. Ventriculomegaly features assessed using our framework’s predictions successfully distinguished NPH from Alzheimer’s disease, post-traumatic volume loss, and headache on a chart-verified VA-Cohort subset (n = 238) with a test-set Area Under the Receiver Operating Characteristic Curve (AUC) of 0.95, closely matching the performance of features assessed using manual AC-PC localization.ConclusionThe proposed registration-guided 3D-UNet framework accurately and automatically localizes the AC-PC on CT despite varied structural degeneration, enabling standardized radiological feature computation. This approach can augment neurodegenerative disease screening on CT, the primary modality for elderly patients evaluated for falls and altered mentation.
IntroductionNeuropsychiatric manifestations, such as impulse control disorders (ICD) and apathy, are common in Parkinson's disease (PD) and significantly impact patient quality of life, yet their underlying neuroanatomical substrates remain poorly understood. This study presents an explainable artificial intelligence (XAI) framework to investigate these phenotypes using structural MRI.MethodsWe employed a 3D Vision Transformer (MobileViT-3D) trained on T1-weighted images from 364 drug-naive PD patients, stratified into four subgroups: idiopathic PD (PD), PD with apathy (PD + A), PD with ICD (PD + ICD), and PD with both conditions (PD + A + ICD). To interpret the model's decision making, Guided Backpropagation and Integrated Gradients were applied to generate saliency maps, followed by a multi-level statistical analysis. Based on these patterns, volumes of interest (VOI) were defined. Conventional morphometric (e.g., cortical thickness) and radiomic texture features were obtained from each VOI applied to each subject.ResultsVoxel-wise analysis of saliency maps suggested group-level differences after correction for multiple comparisons (p ≤ 0.05 FWE) in prefrontal, temporal, and cerebellar regions of the neuropsychiatric conditions compared to PD. While no significant group differences were observed for morphometric features, radiomic texture analysis identified eight features associated with differences between the PD + A + ICD group and idiopathic PD. These features included texture metrics indicative of increased heterogeneity and complexity.DiscussionOverall, these findings provide preliminary evidence that a combined XAI and radiomic approach may help identify subtle, texture-based neuroanatomical signatures associated with the co-occurrence of apathy and ICD in early PD that are not readily captured by conventional univariate analyzes. This framework may support the generation of clinically relevant hypotheses regarding the neural correlates of heterogeneous neuropsychiatric symptoms in PD.
The article presents the Atlantis Source Connectivity Toolbox (ASCT), a pipeline for the complete analysis of source-based directed (effective) connectivity from MEG/EEG signals. Connectivity estimation is implemented with the Directed Transfer Function, a multivariate autoregressive method grounded in the Granger causality approach. The theoretical concepts and challenges in source reconstruction and causal inferences from electrophysiological and biomagnetic recordings are introduced. Importantly, the guidelines for the custom analyses are also presented together with the toolbox user manual. By providing the complete analytic pipeline with practical recommendations, we intend to support standardization and replicability in connectivity studies. The validation of the method is further demonstrated with simulated MEG data, prepared from real resting data recording, as well as with realistic EEG data from the Eriksen flanker task. Our results highlight the importance of routine use of the leakage correction that prevents the appearance of spurious links. We also discuss potential benefits of source separation prior to their reconstruction for more accurate estimates of connectivity.
IntroductionSyncope prediction during head-up tilt testing (HUTT) remains challenging due to the complex interplay between autonomic and cardiovascular responses. This study investigates three computational approaches to forecast HUTT outcomes using continuous electrocardiogram (ECG) and blood pressure recordings from 105 patients with a history of syncope who underwent HUTT following a modified Italian protocol.MethodsBeat-to-beat heart rate and blood pressure signals were analyzed using: (1) gradient boosting models applied to frequency-domain features of heart rate variability (HRV); (2) an analytical modeling approach employing k-nearest neighbors (kNN) regression on transformed physiological signals; and (3) an incremental neural network model.Results and DiscussionAmong these, the kNN regression approach provided the most consistent short-term forecasting of syncope probability, maintaining mean absolute errors below 0.13 for predictions up to 300 s before syncope onset. Gradient boosting models achieved promising classification performance with ROC AUC values up to 0.70, while the incremental network yielded moderate results. These findings demonstrate that data-driven analysis of early physiological changes can enable short-term forecasting of vasovagal syncope during HUTT, supporting the development of predictive tools for clinical risk assessment and personalized syncope management.
The decoding of latent neural states from observable signals is a key focus of modern brain-AI research. Although most neural decoding models are based on electrophysiological recordings, peripheral motor outputs also convey information about circuit-level dynamics. Handwriting is a channeled behavioral signal that indexes the health of the cortico-basal ganglia-thalamo-cortical loop. In Parkinson's disease (PD), dopaminergic loss disrupts this circuit, leading to tremor oscillations, micrographia, and movement irregularities. The challenge of decoding behavior-encoded neural signals from handwriting images constitutes a principled neural signal interpretation problem. We propose a physics-informed and interpretable AI framework for decoding basal ganglia motor dysfunction through harmonic oscillator perturbation analysis. Six energy-inspired measures are extracted to quantify different aspects of motor system dynamics: intensity variation, spatial gradients, multi-scale stability, deviation variability, directional anisotropy, and cross-scale interactions. The transparent mapping between computational models and neurophysiological processes is enabled by these steps, which have their basis in mechanistic theories of amplitude modulation and oscillatory instability. High levels of discrimination were achieved when 594 handwriting trials (279 with PD and 315 controls) were assessed using repeated 10-fold cross-validation for spiral, circle, and meander tasks. Support Vector Machines achieved 84.06% accuracy and 93.56% sensitivity, with highly significant group differences across all features (p < 10-33; Cohen's |d| = 0.87-1.51). Through the integration of physics-based modeling and interpretable machine learning, the proposed framework extends neural signal interpretation beyond direct neural recordings, establishing handwriting as a low-cost, behaviorally encoded biomarker of circuit state and advancing AI-driven decoding of brain dysfunction.
IntroductionAlzheimer's disease (AD) is difficult to treat because of its multifactorial causes and heterogeneous progression across individuals. This study introduces FUSION-AD, a user-friendly and interpretable artificial intelligence framework for AD risk assessment and subgroup discovery.MethodsFUSION-AD integrates tree-based models, transformer-based neural networks, rule mining, and subgroup discovery to provide accurate and interpretable predictions. The framework was developed using the synthetic El Kharoua Alzheimer's Disease Dataset, which contains 2,149 structured clinical records from patients aged 60–90 years, with a mean age of 74.6 years and an average mini-mental state examination (MMSE) score of 21.7. The pipeline included data preprocessing, model benchmarking, feature-importance analysis, SHAP-based explanation, transformer attention analysis, association rule mining, and subgroup discovery.ResultsWithin the evaluated dataset, TabNet achieved the strongest point-estimate performance among the standalone benchmark models on the primary evaluation split, with an area under the receiver operating characteristic curve (AUROC) of 0.95, followed by XGBoost at 0.93, Random Forest at 0.92, and Logistic Regression at 0.89. Feature importance, SHAP values, and transformer attention consistently identified MMSE, Functional Assessment, and Memory Complaints as the most influential predictors. Association rule mining further highlighted diabetes and high body mass index as important risk factors. Subgroup discovery identified four clinical clusters, with prevalence ranging from 21.3 to 28.4%. Cluster 0 showed notable declines in daily functioning, with Functional Assessment decreasing by 2.1 and activities of daily living decreasing by 1.5, whereas Cluster 1 maintained daily functioning but showed increased behavioral symptoms.DiscussionFUSION-AD demonstrates that AD can be modeled in a way that balances predictive performance with interpretability within the studied dataset. The identified subgroup patterns suggest that lifestyle-driven profiles may benefit from preventive strategies, while cognitively impaired groups may require closer monitoring. These findings provide a foundation for future clinically oriented decision-support systems and require further validation using real-world clinical datasets.
A phantom defined by a 3D-printed system of markers representing an anatomical coordinate system is proposed. This phantom is designed for use in magnetic resonance and X-ray imaging devices to validate acquisition parameters and prevent laterality errors during image acquisition and processing. The phantom allows visualization of the coordinate system axes - left-right, posterior-anterior, and inferior-superior - in all orthogonal slices of a volume. A computational method, using the phantom as a reference, is also proposed to automatically detect and correct flips and permutations in RAS coordinate system representations. Testing the phantom across four modalities available in our platforms, followed by conversions of the recordings to the NIfTI format, enables the detection, correction, and adjustment of protocols in 4 out of 5 configurations. The 3D printing models and orientation detection/correction code are shared with the community as open-source and open-access resources to enable quick, cost-effective, and accessible production.
In the era of data-intensive science, managing and verifying research processes that include raw data, analysis scripts, workflows, and documentation, remains a major challenge, particularly in complex fields like neuroimaging were multiple processing and analyses stages can take place. Fragmented tools, inconsistent documentation, and poor platform integration continue to undermine reproducibility. While current Electronic Lab Notebooks (ELNs) aid organization, they often lack cross-platform interoperability, authorship certification, and immutable auditability. To address these gaps, this study proposes a novel ELN framework IntegriLAB, that consolidates traditional data, document and code repositories into a centralized, web-based system for end-to-end project tracking. As a proof of concept, our case study integrates DataLad, Overleaf and GitHub. It allows real-time monitoring across the research cycle while preserving familiar workflows. A key innovation is the use of LabTrace built on green blockchain technology to certify research activities through cryptographically signed, immutable records, ensuring data integrity and verifiable authorship with minimal environmental impact. By unifying project management, secure data handling, and blockchain-based verification, the proposed ELN advances reproducibility, fosters trust, and strengthens collaborative research practices.
Understanding the temporal organization of brain activity requires methods that capture scale-free dynamics while accounting for the high-dimensional, spatially correlated nature of the electroencephalogram (EEG) data. We propose a novel framework that integrates nonlinear manifold learning (Isometric mapping) with detrended fluctuation analysis (DFA) to quantify long-range temporal correlations (LRTC) in the alpha-band of EEG signals. We applied this framework to two music related EEG datasets, as music is known to evoke different emotions and synchronize brain activity. The first dataset was obtained during live Indian classical music (ICM) listening that included two ragas, Yaman and Puriya Dhanashree. EEG was recorded from 13 healthy volunteers (24 channels, sampled at 500 Hz). The second dataset is the Music BCI dataset (006-2015), which includes Jazz and Synth-pop musical clips, with EEG collected from 11 subjects (64 channels, downsampled to 200 Hz). The EEG data from both datasets were preprocessed, band-limited to 8–13 Hz, and segmented into non-overlapping 2-s windows. Alpha-band power was extracted from each channel to form the feature matrix used for embedding. For the ICM dataset, Isometric mapping (Isomap) produced a low-dimensional representation (d = 3), which we analyzed using two approaches: (i) a norm-based approach and (ii) a mean-based approach. For comparison, an equivalent PCA-based pipeline (d = 5) was implemented. The Isomap mean-based DFA yielded consistent scaling exponents (α) in the range of 0.66–0.70, with higher goodness-of-fit (R2) and narrower bootstrap confidence intervals than the norm-based approach. PCA produced similar trends but required more dimensions. Paired t-tests showed that the Isomap mean-based approach detected music-related changes more sensitively than PCA (Yaman p = 0.02; Puriya Dhanashree p = 0.008). Comparable results were also observed for the second Music BCI dataset, where Isomap achieved a compact representation with d = 5, compared to d = 8 for PCA. In this dataset as well, the mean-based DFA yielded α values in the range of 0.62–0.65 and higher goodness-of-fit. Overall, the results suggest that combining nonlinear manifold embeddings with mean-based DFA provides a compact and robust framework for characterizing scale-free temporal structure in EEG data.
Vascular network reconstruction is a crucial step in extracting vessel morphology and establishing its topological relationships from biological imaging data, holding significant scientific importance for studying brain structure and function, metabolism, and disease mechanisms. Current methods for vascular network reconstruction typically follow a “segment-first, then reconstruct” pipeline: first generating a binary segmentation from vascular images, followed by topological modeling. However, due to significant variations in vessel diameters, frequent presence of luminal voids in large vessels, and the complex, densely distributed nature of capillaries, existing approaches still face notable limitations in reconstruction accuracy. To address this, this study introduces and releases an annotated dataset of mouse brain vasculature. The dataset comprises 60 3D image blocks with the size of 512 × 512 × 512 acquired from four mouse brain samples using fluorescence micro-optical sectioning tomography (fMOST) imaging. It encompasses diverse structural morphologies ranging from large vessels to capillaries, with detailed annotations specifically targeting challenging vascular regions. Additionally, we provide a standardized vascular annotation pipeline and associated tools. This dataset aims to serve as a benchmark to support the development, evaluation, and comparison of algorithms for vascular network segmentation, reconstruction, and related tasks.
Reliable imaging biomarkers are essential for improving early detection of Alzheimer’s disease (AD). We evaluated whether an automated MRI-based classifier provides diagnostic performance comparable to expert radiologists in differentiating cognitively normal (CN) individuals from patients with AD using standardized ADNI data. Thirty-eight structural MRI datasets (20 CN, 18 AD) were analyzed. An automated multi-class volumetric classifier and two board-certified radiologists independently assigned probability scores across seven diagnostic categories. Performance was evaluated using a partial-credit scoring rule to account for probabilistic ties. Diagnostic performance for CN-AD discrimination was assessed using accuracy, sensitivity, specificity, receiver operating characteristic (ROC) analysis, inter-observer agreement metrics, Brier scores for calibration, and decision curve analysis (DCA) for clinical utility. The automated classifier achieved an accuracy of 0.66, sensitivity of 0.56, and specificity of 0.75. Radiologists demonstrated comparable performance with inherent inter-observer variability. Agreement between automated and human assessments was fair at the categorical level, with low concordance for continuous probability estimates. ROC analysis based on continuous AD probabilities demonstrated high discrimination performance for the automated model (AUC = 0.90), exceeding that of radiologists (AUC = 0.71 and 0.62). DCA indicated that the automated pipeline provides a positive net benefit as a second-opinion tool. This exploratory study emphasizes the impact of evaluation frameworks on performance metrics and supports further validation using multi-modal data in larger cohorts.
BackgroundBrain tumor diagnosis from magnetic resonance imaging (MRI) remains a challenging task due to the high variability in tumor appearance and the limitations of manual interpretation.MethodsTo address these challenges, this paper proposes NeuroFusionNet, a deep learning framework for automated brain tumor classification from MRI. The framework integrates GAN-based synthetic image generation with transfer learning using a fine-tuned VGG16 backbone. Real and GAN-generated MRI images are passed through VGG16 to extract discriminative feature representations, which are then used for final classification. To adapt the model to domain-specific MRI characteristics while preserving pretrained knowledge, the last ten layers of VGG16 are fine-tuned and the remaining layers are kept frozen.ResultsThe effectiveness of NeuroFusionNet is validated on two publicly available brain MRI datasets. Experimental results demonstrate that the proposed learning framework achieves classification accuracies of 99.05 and 98.75% on the Brain Tumor MRI Dataset and the MRI with Bounding Boxes Dataset, respectively, consistently outperforming several state-of-the-art neural architectures, including VGG16, VGG19, MobileNetV2, DenseNet121, and NASNetLarge.ConclusionThe results suggest that NeuroFusionNet is effective for the evaluated public MRI datasets; additional external validation is required.
Neuroimaging presents us with an in-depth understanding about brain structure and function, yet the data complexity poses significant analytical challenges. Current frameworks suffer from issues such as scalability, poor integration with traditional statistics and a need for a programing background, which hinder researchers from focusing on neuroscience questions. To address these limitations, we present BrainInsights, an integrated and automated GUI-based pipeline ecosystem designed to facilitate the analysis of multi-modal or multi-parametric neuroimaging data in a flexible way. The framework comprises three core tools: MARIA (MAgnetic Resonance Imaging data Analysis and inspection tool) for data inspection and hypotheses testing, ML Pipeline for automated feature selection and model construction, and ML DaViz for model evaluation and bio-signature generation. Deployed as a singularity container, the system ensures reproducibility and scalability across computing environments. We validated BrainInsights using diverse datasets, including multi-parametric MRI studies of Anorexia Nervosa, Crohn’s disease, and Rheumatoid Arthritis. Specifically, the framework distinguished young Anorexia Nervosa patients from controls with a balanced accuracy of 65%, while in the PreCePRA trial, it predicted Rheumatoid Arthritis treatment response with a balanced accuracy of up to 95.4% using functional pain markers. The results demonstrate the ability of the framework to achieve high separation of subgroups and treatment success and additionally bridge hypotheses-driven statistical analysis with data-driven machine learning analysis. By enabling interpretability tools like SHAP, BrainInsights empowers researchers to move beyond “black-box” modeling to uncover stable, biologically plausible bio-signatures. Ultimately, this framework aids in accelerating the translation of complex neuroimaging data into meaningful clinical insights.
Purpose:Dyslexia is a prevalent neurodevelopmental disorder that impairs a children's ability to reading, writing, and language processing despite normal cognitive skills. Early identification is vital for timely support and interventions in children with dyslexia. This study aimed to develop an efficient EEG-based pipeline for dyslexia detection using deep learning techniques, while providing a consistent evaluation protocol for fair comparison across models and prior approaches. Methods:EEG recordings were acquired from 51 participants (26: dyslexic and 25: non-dyslexic), aged 5-10 years, during cognitive task performance. These signals were processed, segmented, and decomposed into standard frequency bands (alpha, beta, delta, and theta) using the discrete wavelet transform to capture discriminative neural patterns. Filter-based feature selection techniques were applied before classification to optimize performance and reduce redundancy to identify the most informative features. These ranked and individual band-wise features were systematically evaluated with classical machine learning baselines (Decision Trees, SVM, k-NN, and ensemble learners) alongside the proposed deep neural networks. In addition, we benchmarked end-to-end raw-EEG deep learning baselines (1D-CNN, LSTM, and EEGNet) and re-implemented representative existing pipelines, all evaluated on our dataset using the same evaluation protocol. Results:The proposed compact deep neural network with four hidden layers achieved the best performance, reaching classification accuracy of 98.85%, outperforming all baseline models, raw-EEG deep learning baselines, and re-implemented approaches. Conclusion:These findings support the feasibility of DWT-driven EEG analysis combined with deep learning for more accurate and early dyslexia detection. The proposed approach holds promise as a non-invasive screening tool to support improved educational outcomes through early diagnosis and targeted intervention.
Background:Secondary quantitative analysis of brain magnetic resonance imaging (MRI) can provide valuable information for many neurological diseases, including multiple sclerosis (MS), but it demands complete datasets that are often unavailable clinically. We investigated how image synthesis via deep learning using cycle-consistent generative adversarial networks (CycleGANs) compared with Pix2Pix as a related method, based on T1-weighted and T2-weighted brain MRI in MS, following verification on two streamlined datasets. The synthesized images were also evaluated against the source data. Methods:The streamlined datasets involved 1,113 healthy participants from the Human Connectome Project (HCP) and 318 participants from the Parkinson's Progression Markers Initiative (PPMI). The MS cohort in this study included 105 participants scanned with different protocols. Image synthesis was bidirectional between T1- and T2-weighted MRI using CycleGAN with and without spectral normalization, as well as Pix2Pix. Utility testing focused on T1-weighted MRI that was most often unavailable in MS, and that involved lesion detection, brain volumetry, and lesion texture analysis. Results:All CycleGAN models performed competitively, while Pix2Pix performed better, mostly with streamlined datasets (p < 0.001). The average peak signal-to-noise ratio ranged from 24.860-28.570 versus 28.520-31.100, and the structural similarity index ranged from 0.838-0.901 versus 0.924-0.943. With spectral normalization, CycleGAN improved in PPMI but not in HCP and generally not in MS (p < 0.001). Furthermore, the synthesized images showed high similarity to the source data in utility tests, although Pix2Pix T1 images appeared more heterogeneous in lesion texture than source T1 images. Conclusion:CycleGAN without spectral normalization appeared feasible for synthesizing common clinical brain MRI, including T1-weighted images usable for subsequent quantitative analysis in MS.
