Learnable Diffusion Framework for Mouse V1 Neural Decoding

Area of Science:

• Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Decoding neural signals is key to understanding brain representations.
• Human fMRI studies dominate visual stimulus reconstruction.
• Mouse calcium imaging offers single-neuron resolution in non-primate visual systems.

Purpose of the Study:

• Introduce Sensorium-Viz, a diffusion-based framework for decoding mouse primary visual cortex activity.
• Enable high-resolution image reconstruction from single-neuron responses.
• Provide insights into non-primate visual processing.

Main Methods:

• Developed Sensorium-Viz, integrating a Diffusion Transformer (DiT) with a spatial neuron-embedding module.
• Implemented a synthetic-response augmentation strategy for improved performance and generalization.
• Utilized calcium imaging data from the mouse primary visual cortex.

Main Results:

• Sensorium-Viz reliably reconstructs complex, high-resolution images from novel single-neuron data.
• Synthetic-response augmentation improved reconstruction by over 30% and enabled cross-mouse generalization.
• Achieved up to 10.65% performance gain over leading fMRI-based methods.

Conclusions:

• Sensorium-Viz is a robust tool for vision decoding in mouse models.
• Neurons sensitive to low-level visual features are fundamental to V1's stimulus representation.
• The framework advances single-neuron-level understanding of non-primate visual systems.