Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

FreqSelect: Frequency-Aware fMRI-to-Image Reconstruction

Created by
  • Haebom

Author

Junliang Ye, Lei Wang, Md Zakir Hossain

Outline

This paper addresses the task of reconstructing natural images from fMRI data. Existing two-stage models (combining VAE and diffusion models) suffer from inefficient processing of all spatial frequency components equally. In this paper, we propose FreqSelect, a lightweight, adaptive module that selectively filters spatial frequency bands. FreqSelect acts as a content-aware gate between image features and natural data by highlighting the most relevant frequencies for brain activity prediction and suppressing irrelevant frequencies. Evaluation on the Natural Scenes dataset demonstrates improved reconstruction quality in both low- and high-level metrics, and the learned frequency selection patterns provide interpretable insights into how various visual frequencies are represented in the brain. Furthermore, the proposed approach generalizes across subjects and scenes and has the potential to be extended to other neuroimaging techniques.

Takeaways, Limitations

•
Takeaways:
◦
Improving the accuracy of natural image reconstruction from fMRI data.
◦
We observed performance improvements across both low-level and high-level metrics.
◦
Learned frequency selection patterns provide interpretable insights into visual frequency representations in the brain.
◦
Excellent generalization performance across topics and scenes.
◦
Potential extension to other neuroimaging techniques.
•
Limitations:
◦
Evaluation was conducted only on the Natural Scenes dataset. Generalization performance on other datasets is required.
◦
Further studies are needed to determine the biological relevance of the FreqSelect module.
◦
Although the possibility of extension to other neuroimaging techniques has been suggested, further research is required for practical application and performance verification.
👍
Made with Slashpage