Topological Receptive Field Model for Human Retinotopic Mapping

Yanshuai Tu, Duyan Ta, Zhonglin Lu, Yalin WangMICCAI 2021 · OpenReview

TL;DR

This paper introduces a novel "topological receptive field" (tRF) model to improve the accuracy and biological plausibility of human retinotopic maps derived from fMRI data. By enforcing topological consistency during the decoding process, the tRF model significantly reduces errors and generates more reliable maps compared to existing methods.

What Problem Does This Paper Address?

Understanding how the brain represents visual information is a fundamental goal in neuroscience. Retinotopic maps, which describe how visual input from the retina is organized in the visual cortex, are crucial for this understanding. These maps are typically generated using functional magnetic resonance imaging (fMRI), which measures brain activity. However, standard fMRI analysis methods often produce "non-topological" results. This means that the spatial relationships between receptive fields (the areas of visual space that neurons respond to) in the brain don't accurately reflect the spatial relationships in the visual input. This is like trying to draw a map where nearby locations on the map are actually far apart in reality. This lack of topological consistency hinders our ability to accurately simulate visual perception, understand brain plasticity, and even diagnose visual impairments.

Methods

The core innovation of this paper is the topological receptive field (tRF) model. Unlike traditional methods that analyze each brain voxel (3D pixel) independently, the tRF model explicitly incorporates the topological constraint. This is achieved through a two-pronged approach:

  1. Topology-Preserving Segmentation: The model first segments the visual cortex into distinct visual areas (like V1, V2, etc.). This segmentation is guided by a "topology-prior" and uses a diffeomorphic registration technique to ensure that the segmentation respects the known topological organization of these areas across individuals. This is crucial because the precise boundaries of visual areas are often determined by the retinotopic mapping itself, creating a circular dependency that the tRF model addresses.
  2. Topological fMRI Decoding: Within each segmented visual area, the tRF model decodes the fMRI signals to estimate the parameters of the population receptive field (pRF) model. Critically, this decoding process is constrained by the topological condition, meaning that neighboring neurons in the cortex are forced to have receptive fields that are also spatially close in the visual field. This is achieved by incorporating a "Beltrami coefficient" into the optimization problem, which quantifies deviations from conformality (a measure of how well a mapping preserves local angles and shapes, crucial for topological consistency).

The authors also developed an efficient numerical method to solve these coupled problems iteratively, alternating between refining the segmentation and updating the retinotopic map parameters until convergence.

Datasets

The authors evaluated their tRF model on two types of data:

  • Synthetic Data: They generated synthetic fMRI data based on a "double-sech" model, which is a common ground truth for retinotopic mapping studies. This allowed for a direct comparison of their method against ground truth and other established methods. They introduced varying levels of noise to simulate realistic fMRI data.
  • Real Data: They used data from the Human Connectome Project (HCP) dataset, which provides high-resolution 7T fMRI scans with accompanying retinotopy experiments. This dataset is publicly available and represents a valuable benchmark for retinotopic mapping research.

Key Insights and Ideas

  • Topology is Paramount: The paper strongly argues that topological consistency is not just a desirable feature but an essential requirement for biologically plausible retinotopic maps. Ignoring it leads to significant errors.
  • Decoding-Time Constraint is Key: Instead of trying to fix topological violations after decoding (as in many previous methods), the tRF model enforces topological constraints during the decoding process. This is a more principled approach.
  • Segmentation and Decoding are Intertwined: The tRF model recognizes that accurate visual area segmentation and retinotopic map decoding are mutually dependent. Their iterative approach effectively refines both simultaneously.
  • Generalizability: The authors suggest that the tRF framework is general and can be extended to other spatially organized sensory maps, such as auditory maps.

Key Results

The tRF model demonstrated superior performance compared to several established methods, including the standard pRF model, model-fitting, smoothing, and registration-based approaches:

  • Synthetic Data: The tRF model achieved zero topological violations (Tn=0) while maintaining competitive precision in estimating receptive field centers (|Δv|) and sizes (|Δσ|). Other methods, while sometimes achieving good precision, often resulted in a significant number of topological errors.
  • Real Data (HCP): For the HCP dataset, the tRF model achieved the best fMRI fitting goodness (lowest RMSE of fitting error) and, crucially, zero topology violations. For example, for one observer, the tRF model resulted in Tn=0, while the pRF model had Tn=870, and other methods also showed substantial violations. The RMSE of fitting error was also the lowest with the tRF model.

Limitations and Future Work

  • Computational Cost: While the authors developed an efficient scheme, the iterative nature of the tRF model likely still incurs a higher computational cost than simpler, non-topological methods.
  • Generalizability to Other Sensory Maps: While suggested, the paper primarily focuses on retinotopic mapping. Further validation on other sensory modalities would be beneficial.
  • Assumptions of Topology-Prior: The segmentation relies on a "topology-prior." The robustness of the method to inaccuracies or variations in this prior could be explored.

The authors suggest that the framework's generality makes it applicable to other sensory maps, indicating a promising avenue for future research in understanding various brain representations.

Summary updated

Abstract

The mapping between visual inputs on the retina and neuronal activations in the visual cortex, i.e., retinotopic map, is an essential topic in vision science and neuroscience. Human retinotopic maps can be revealed by analyzing the functional magnetic resonance imaging (fMRI) signal responses to designed visual stimuli in vivo. Neurophysiology studies summarized that visual areas are topological (i.e., nearby neurons have receptive fields at nearby locations in the image). However, conventional fMRI-based analyses frequently generate non-topological results because they process fMRI signals on a voxel-wise basis, without considering the neighbor relations on the surface. Here we propose a topological receptive field (tRF) model which imposes the topological condition when decoding retinotopic fMRI signals. More specifically, we parametrized the cortical surface to a unit disk, characterized the topological condition by tRF, and employed an efficient scheme to solve the tRF model. We tested our framework on both synthetic and human fMRI data. Experimental results showed that the tRF model could remove the topological violations, improve model explaining power, and generate biologically plausible retinotopic maps. The proposed framework is general and can be applied to other sensory maps.
BibTeX
@inproceedings{tu2021topological,
  author = {Yanshuai Tu and Duyan Ta and Zhonglin Lu and Yalin Wang},
  title = {{Topological Receptive Field Model for Human Retinotopic Mapping}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2021},
  url = {https://openreview.net/forum?id=febKfbjSbFs},
}
Figure