Digital 3D atlases of the human brain are fundamental tools to understand its functionally relevant topography. Most atlases today provide a resolution in the millimeter range, allowing to study structure-function relationships of areas and large networks, but failing to integrate information about thin fiber bundles, cortical layers, columns, microcircuits or cells. To understand brain organization, we have to bridge the spatial scales and develop models at the level of 1-20 micrometers, which is beyond most advanced Highfield MRI techniques.
Over the last decade, AI has come to be dominated by ML, and in particular DL. DL is often associated with particular tasks (e.g. categorizing images), but it is, in fact, a very general program of research. This general program is defined by two principles: (1) it is best to avoid hard-wiring any function into an AI, i.e. as much should be learned as possible, (2) learning should be done using hierarchical artificial neural networks (ANN) that are optimized in an end-to-end manner, i.e. all components of the hierarchy should be altered by learning. However, despite the first principle that hard-wiring should be avoided, it cannot be completely avoided. The reason is the classic bias-variance tradeoff from statistics: any model is faced with the dilemma that avoiding any bias altogether typically leads to unacceptable levels of variance in outcome. In ANNs, this means that a network with no built-in assumptions about the task to be learned will be overly complex, and prone to either overfitting or slow convergence (or both). Thus, most DNNs actually do utilize some built-in “inductive biases” to help minimize variance. Given the success of stochastic gradient descent and its offshoots, a significant portion of modern Machine Learning research is devoted not to crafting new learning algorithms, but rather, to crafting networks with architectures that ensure useful inductive biases.
The Goals of this Research Topic are:
1. Achieve a full coverage parcellation of BigBrain’s cortex into areas, layers and subcortical structures, and generate a second high-resolution model which goes down to the level of individual cells (“Cellular BigBrain”), both constituting a
novel reference framework at unmatched level of detail that is freely available and follows FAIR data sharing principles.
2. Extend the computational model by high-resolution 3D measurements of structural connectivity as provided by 3D Polarized Light Imaging (PLI), as well as 3D reconstruction of receptor autoradiography, providing a deep characterization of the brain’s molecular organization.
3. Develop novel AI methods for 2D and 3D biomedical image segmentation and classification, that can deal with very limited amounts of training data by incorporating basic topographical rules, and be applied to large data volumes by leveraging HPC.
4. Integrate the neuroimaging, brain network simulation and AI research.
This Research Topic in “Computational Intelligence for Integrative Neuroscience Through High-Performance Neuroimaging Data Analysis” welcomes submissions of Original Research, Review, Methods, and Perspective articles focused on the following sub-topics:
- Brain parcellation using AI-based segmentation
- Mapping cytoarchitectonic areas in cerebral cortex is central to creating multi-modal multi-level
brain atlases and a gold standard for describing its cellular structure.
- Multimodal data integration with AI.
- Cross-modal, cross-subject mapping of histological data to in vivo MRI with Computational Intelligence.
- Computational Intelligence for 7T in vivo MRI.
- 20 µm reconstruction of a new brain sample with AI
- Develop Parallel & Scalable Workflows for Deep Learning
- Exploit Maximum Data Accessibility for Neuroscience Datasets
- Brain-Inspired Artificial Intelligence
Digital 3D atlases of the human brain are fundamental tools to understand its functionally relevant topography. Most atlases today provide a resolution in the millimeter range, allowing to study structure-function relationships of areas and large networks, but failing to integrate information about thin fiber bundles, cortical layers, columns, microcircuits or cells. To understand brain organization, we have to bridge the spatial scales and develop models at the level of 1-20 micrometers, which is beyond most advanced Highfield MRI techniques.
Over the last decade, AI has come to be dominated by ML, and in particular DL. DL is often associated with particular tasks (e.g. categorizing images), but it is, in fact, a very general program of research. This general program is defined by two principles: (1) it is best to avoid hard-wiring any function into an AI, i.e. as much should be learned as possible, (2) learning should be done using hierarchical artificial neural networks (ANN) that are optimized in an end-to-end manner, i.e. all components of the hierarchy should be altered by learning. However, despite the first principle that hard-wiring should be avoided, it cannot be completely avoided. The reason is the classic bias-variance tradeoff from statistics: any model is faced with the dilemma that avoiding any bias altogether typically leads to unacceptable levels of variance in outcome. In ANNs, this means that a network with no built-in assumptions about the task to be learned will be overly complex, and prone to either overfitting or slow convergence (or both). Thus, most DNNs actually do utilize some built-in “inductive biases” to help minimize variance. Given the success of stochastic gradient descent and its offshoots, a significant portion of modern Machine Learning research is devoted not to crafting new learning algorithms, but rather, to crafting networks with architectures that ensure useful inductive biases.
The Goals of this Research Topic are:
1. Achieve a full coverage parcellation of BigBrain’s cortex into areas, layers and subcortical structures, and generate a second high-resolution model which goes down to the level of individual cells (“Cellular BigBrain”), both constituting a
novel reference framework at unmatched level of detail that is freely available and follows FAIR data sharing principles.
2. Extend the computational model by high-resolution 3D measurements of structural connectivity as provided by 3D Polarized Light Imaging (PLI), as well as 3D reconstruction of receptor autoradiography, providing a deep characterization of the brain’s molecular organization.
3. Develop novel AI methods for 2D and 3D biomedical image segmentation and classification, that can deal with very limited amounts of training data by incorporating basic topographical rules, and be applied to large data volumes by leveraging HPC.
4. Integrate the neuroimaging, brain network simulation and AI research.
This Research Topic in “Computational Intelligence for Integrative Neuroscience Through High-Performance Neuroimaging Data Analysis” welcomes submissions of Original Research, Review, Methods, and Perspective articles focused on the following sub-topics:
- Brain parcellation using AI-based segmentation
- Mapping cytoarchitectonic areas in cerebral cortex is central to creating multi-modal multi-level
brain atlases and a gold standard for describing its cellular structure.
- Multimodal data integration with AI.
- Cross-modal, cross-subject mapping of histological data to in vivo MRI with Computational Intelligence.
- Computational Intelligence for 7T in vivo MRI.
- 20 µm reconstruction of a new brain sample with AI
- Develop Parallel & Scalable Workflows for Deep Learning
- Exploit Maximum Data Accessibility for Neuroscience Datasets
- Brain-Inspired Artificial Intelligence
