Neural Noise

Deep learning with spiking neural networks for multimodal neuroimaging

2 min read

  • A review covers deep learning models using spiking neural networks for modeling and analysis of multimodal neuroimaging data.1
  • SNNs bridge computational neuroscience and neural data analysis and are relevant for BCI decoding and neuroinformatics.1 1

Weekly enrichment (2026-07-20)

  • The underlying article is a review in Frontiers in Neuroscience (2025, doi 10.3389/fnins.2025.1623497) on deep-learning models built with spiking neural networks (SNNs) for modeling and analyzing multimodal neuroimaging data.2
  • The authors ran a quantitative and thematic analysis of 21 selected publications to characterize research trends, topics, and geographical contributions in the SNN-neuroimaging field.2
  • The review reports that SNNs outperform traditional deep-learning models (such as CNNs) on classification, feature extraction, and prediction in terms of precision, recall, and accuracy—especially when fusing multiple modalities—while consuming less energy than conventional networks.2
  • SNNs are being applied across neuroimaging modalities including structural MRI, functional MRI (fMRI), electroencephalography (EEG), and diffusion tensor imaging (DTI), enabling personalized, dynamic modeling of brain function.2
  • A NeuCube 3D brain-template SNN using spike-timing-dependent plasticity (STDP) learning classified and predicted cognitive decline—dementia and mild cognitive impairment—two years ahead with reported accuracies of 91% and 95%.2
  • In EEG applications, NeuCube reached 97% classification accuracy distinguishing epilepsy, migraine, and healthy subjects, surpassing bidirectional LSTM (90%) and reservoir-based SNN (85%) while requiring fewer training samples.3
  • A NeuCube-based framework fusing multimodal MRI data (T1-weighted, T2-FLAIR, DTI, and fMRI) achieved 88% accuracy for Alzheimer’s disease detection and is efficiently implementable on low-power neuromorphic hardware for real-time use.4
  • The reviews consistently flag persistent obstacles—multimodal data fusion, high computational/training demands, scarce large-scale datasets, ANN-to-SNN conversion inefficiencies, and a narrow focus on classification—positioning SNNs as an energy-efficient, biologically plausible bridge between neuroscience and AI relevant to BCI decoding and neuroinformatics.2 5

Footnotes

  1. https://news.google.com/rss/articles/CBMilAFBVV95cUxQWVItNXdhRVBDckZnQkJ5cDZjUV9yWkU1elhoNVlialJqc0hTdVRUSUpBSlZITE1mbHloR29CMHJpTjZUTzVuMThTMGtlZ0MxNWhZcm9weUNHcXhTOVRaWWdzZGZ5WHFEOERSZzQzbmFMQ0trNFFsbnVPRWRvcGQ3cnRCM3dPY0RnNGlOal92ZHZreGFr?oc=5 ↩ ↩2 ↩3

  2. https://doi.org/10.3389/fnins.2025.1623497 ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  3. https://doi.org/10.3390/bioengineering12060628 ↩

  4. https://doi.org/10.36227/techrxiv.176281061.13895617/v1 ↩

  5. https://www.mdpi.com/1424-8220/25/21/6747 ↩

Graph View