Rafat Damseh

Objective. Diffusion magnetic resonance imaging (MRI) signals are sensitive to microvascular geometry and hemodynamics, but current approaches rely on computationally intensive simulations on optical coherence tomography (OCT)-derived vascular graphs. We investigated whether three-dimensional deep learning models can serve as fast surrogates that predict diffusion MRI responses directly from microvascular OCT angiograms. Approach. Five mouse cortical OCT angiograms were segmented into 4792 vascular subgraphs and converted into nine-channel image volumes encoding vascular geometry, magnetic field gradient, hematocrit, oxygen saturation, partial oxygen pressure, blood velocity magnitude, and three velocity components. For each sample, a physics-based framework generated spin echo (SE) and diffusion-weighted signals for three b -values (50, 100, 500 s mm − 2 ), three gradient timing pairs, and ten gradient directions. We trained five 3D architectures (AutoEncoder, BasicUNet, DenseNet169, EfficientNetB4 and ResNet) with late fusion of acquisition parameters and evaluated performance using MSE, MAE and R 2 . Main results. Under within-distribution (random-split) evaluation, test R 2 values were at least 0.87. DenseNet169 achieved the best overall accuracy (MSE = 0.0039 ± 0.0029 , MAE = 0.0397 ± 0.0161 , R 2 = 0.92 ) and maintained R 2 > 0.80 across all diffusion settings, while ResNet provided near-perfect reconstruction of the SE sequence alone ( R 2 ≈ 1.00 ). Leave-one-subject-out cross-validation gave MSE = 0.0173 ± 0.0012 , MAE = 0.094 ± 0.004 , and R 2 = 0.666 ± 0.019 , with 34.4% of held-out samples reaching R 2 ⩾ 0.80 . Channel ablation revealed velocity magnitude and velocity components as dominant predictors. Once trained, the models predicted full diffusion responses in milliseconds, providing speedups of four to six orders of magnitude compared with simulations. Significance. These results demonstrate that deep learning surrogates can accurately and efficiently approximate diffusion MRI responses from microvascular OCT angiograms, enabling rapid exploration of microvascular configurations and acquisition protocols.

Background: Epilepsy affects approximately 50 million people globally and imposes a substantial clinical and societal burden, requiring continuous and personalized monitoring for effective management. Wearable artificial intelligence (AI) technologies offer a promising solution by leveraging physiological signals and machine learning for seizure detection and prediction. While various approaches have been proposed, a comprehensive overview summarizing these advances and challenges is still needed. Objective: This review aims to comprehensively explore and map the existing literature on AI-driven wearable technologies for epilepsy, identifying device characteristics, AI methodologies, biosignal measurements, validation approaches, and research gaps. Methods: A scoping review was conducted following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) guidelines. A systematic search was performed across six electronic databases (Scopus, MEDLINE, Embase, ACM Digital Library, IEEE Xplore, and Google Scholar) to identify relevant studies published up to December 2023. We included studies that developed AI algorithms for epilepsy using noninvasive wearable devices (eg, smartwatches, smart clothing) and excluded those using nonwearables or in-body devices. Eligible publication types included journal articles, conference papers, and dissertations. Study selection and data extraction were performed independently by six reviewers. The extracted data were synthesized narratively. Results: A total of 67 studies met the inclusion criteria. Research in this domain has increased significantly since 2021, with India, the United States, and China leading contributions. The studies examined both commercial (n=31, 46.3%) and noncommercial (n=31, 46.3%) wearable devices, with Empatica smart bands being the most frequently used. The primary biosignals monitored included activity measures (n=36, 53.7%), cardiovascular metrics (n=33, 49.3%), brain activity (n=24, 35.8%), and skin conductance (n=23, 34.3%). The most common AI models were support vector machines (n=28, 41.8%), random forests (n=14, 20.9%), and convolutional neural networks (n=10, 14.9%). Most models focused on seizure detection (n=54, 80.6%) compared to seizure prediction (n=14, 20.9%), reflecting a research imbalance that suggests the need for further development in predictive analytics. Sensitivity (n=54, 80.6%) was the most frequently reported performance metric, indicating a focus on identifying seizures; however, comprehensive clinical validation remains limited. Closed-source data predominated (n=44, 65.7%), limiting the generalizability of findings. The most used validation methods were leave-one-out cross-validation (n=21, 31.3%) and k-fold cross-validation (n=20, 29.9%), while video electroencephalography served as the primary reference standard (n=35, 52.2%). Conclusions: Wearable AI technologies show significant promise in epilepsy management, offering real-time, continuous monitoring and early seizure detection. To realize clinical impact, future research should prioritize the standardization of validation methods, promote open data exchange for reproducibility, and develop energy-efficient algorithms that support real-world deployment in wearable devices.

Modeling microscopic cerebrovascular networks is essential for understanding cerebral blood flow and oxygen transport. High-resolution imaging modalities, such as Optical Coherence Tomography (OCT) and Two-Photon Microscopy (TPM), are widely used to capture microvascular structure and topology. Although TPM angiography generally provides better localization and image quality than OCT, its use is impractical in studies involving fluorescent dye leakage. Here, we exploit generative adversarial learning to produce high-quality TPM angiographies from OCT vascular stacks. We investigate the use of 2D and 3D cycle generative adversarial networks (CycleGANs) trained on unpaired image samples. We evaluate the generated TPM vascular structures based on image similarity and signal-to-noise ratio. Additionally, we evaluated the generated vascular structures after applying vessel segmentation and extracting their 3D topological models. Our results demonstrate that the 2D adversarial learning model outperforms the 3D model in terms of image quality. However, our statistical comparisons of vascular network features show the 3D model’s consistent superiority in generating vascular structures. Our work provides a complementary approach to enhance vascular analysis when only OCT imaging is available.

Background: Neuroimaging segmentation is increasingly important for diagnosing and planning treatments for neurological diseases. Manual segmentation is time-consuming, apart from being prone to human error and variability. Transformers are a promising deep learning approach for automated medical image segmentation. Objective: This scoping review will synthesize current literature and assess the use of various transformer models for neuroimaging segmentation. Methods: A systematic search in major databases, including Scopus, IEEE Xplore, PubMed, and ACM Digital Library, was carried out for studies applying transformers to neuroimaging segmentation problems from 2019 through 2023. The inclusion criteria allow only for peer-reviewed journal papers and conference papers focused on transformer-based segmentation of human brain imaging data. Excluded are the studies dealing with nonneuroimaging data or raw brain signals and electroencephalogram data. Data extraction was performed to identify key study details, including image modalities, datasets, neurological conditions, transformer models, and evaluation metrics. Results were synthesized using a narrative approach. Results: Of the 1246 publications identified, 67 (5.38%) met the inclusion criteria. Half of all included studies were published in 2022, and more than two-thirds used transformers for segmenting brain tumors. The most common imaging modality was magnetic resonance imaging (n=59, 88.06%), while the most frequently used dataset was brain tumor segmentation dataset (n=39, 58.21%). 3D transformer models (n=42, 62.69%) were more prevalent than their 2D counterparts. The most developed were those of hybrid convolutional neural network-transformer architectures (n=57, 85.07%), where the vision transformer is the most frequently used type of transformer (n=37, 55.22%). The most frequent evaluation metric was the Dice score (n=63, 94.03%). Studies generally reported increased segmentation accuracy and the ability to model both local and global features in brain images. Conclusions: This review represents the recent increase in the adoption of transformers for neuroimaging segmentation, particularly for brain tumor detection. Currently, hybrid convolutional neural network-transformer architectures achieve state-of-the-art performances on benchmark datasets over standalone models. Nevertheless, their applicability remains highly limited by high computational costs and potential overfitting on small datasets. The heavy reliance of the field on the brain tumor segmentation dataset hints at the use of a more diverse set of datasets to validate the performances of models on a variety of neurological diseases. Further research is needed to define the optimal transformer architectures and training methods for clinical applications. Continuing development may make transformers the state-of-the-art for fast, accurate, and reliable brain magnetic resonance imaging segmentation, which could lead to improved clinical tools for diagnosing and evaluating neurological disorders.

Epilepsy has been analyzed through uni-modality non-invasive brain measurements such as electroencephalogram (EEG) signal, but identifying seizure patterns is more challenging due to the non-stationary nature of the brain activity and various non-brain artifacts. In this paper, we leverage a vision transformer model (ViT) to classify three types of seizure patterns based on multimodal EEG and functional near-infrared spectroscopy (fNIRS) recordings. We used spectral encoding techniques to capture temporal and spatial relationships for brain signals as feature map inputs to the transformer architecture. We evaluated model performance using the receiver operating characteristic (ROC) curves and the area under the curve (AUC), demonstrating that multimodal EEG-fNIRS signals improved the classification accuracy of seizure patterns. Our work showed that power spectral density (PSD) features often led to better results than features derived from dynamic mode decomposition (DMD), particularly for seizures with high-frequency oscillations (HFO) and generalized spike-and-wave discharge (GSWD) patterns, with an accuracy of 93.14% and 91.69%, respectively. Low-voltage fast activity (LVFA) seizures achieved consistently high performance in EEG, fNIRS, and multimodal EEG-fNIRS setups. Overall, our findings suggest the effectiveness of using the ViT architecture with multimodal brain data accompanied by appropriate spectral features to classify the neural activity of epileptic seizure patterns.

In the complex and multidimensional field of medicine, multimodal data are prevalent and crucial for informed clinical decisions. Multimodal data span a broad spectrum of data types, including medical images (eg, MRI and CT scans), time-series data (eg, sensor data from wearable devices and electronic health records), audio recordings (eg, heart and respiratory sounds and patient interviews), text (eg, clinical notes and research articles), videos (eg, surgical procedures), and omics data (eg, genomics and proteomics). While advancements in large language models (LLMs) have enabled new applications for knowledge retrieval and processing in the medical field, most LLMs remain limited to processing unimodal data, typically text-based content, and often overlook the importance of integrating the diverse data modalities encountered in clinical practice. This paper aims to present a detailed, practical, and solution-oriented perspective on the use of multimodal LLMs (M-LLMs) in the medical field. Our investigation spanned M-LLM foundational principles, current and potential applications, technical and ethical challenges, and future research directions. By connecting these elements, we aimed to provide a comprehensive framework that links diverse aspects of M-LLMs, offering a unified vision for their future in health care. This approach aims to guide both future research and practical implementations of M-LLMs in health care, positioning them as a paradigm shift toward integrated, multimodal data–driven medical practice. We anticipate that this work will spark further discussion and inspire the development of innovative approaches in the next generation of medical M-LLM systems.