Paper Abstract

Understanding how intrinsic brain activity constrains stimulus-evoked responses is a central challenge in cognitive neuroscience. Task-based fMRI measures localized evoked activation but is costly and difficult to scale, whereas resting-state fMRI (rs-fMRI) is easier to acquire yet lacks direct stimulus interpretability. We study resting-conditioned synthesis of visually evoked fMRI (ve-fMRI): given an rs-derived intrinsic functional prior and a visual stimulus image, we predict the corresponding trial-wise activation volume. We propose Rest2Visual, a 3D encoder–decoder network that encodes the resting prior into hierarchical volumetric features and injects stimulus embeddings into the decoder via adaptive group normalization to produce stimulus-specific maps. Using large-scale rs–stimulus–evoked triplets, Rest2Visual improves voxel-wise and representational agreement over stimulus-agnostic and modality-ablated baselines, and preserves subject-specific response structure. We further use image reconstruction as a downstream probe to assess whether the synthesized activations retain stimulus information. The results suggest that individualized intrinsic organization can be transformed into stimulus-aligned activation patterns for scalable functional modeling.

My Contributions

My primary contributions to the Rest2Visual project were focused on the two critical ends of the model pipeline: data preparation and downstream validation.

At the input end, I was responsible for curating and preprocessing the large-scale dataset required to train the 3D encoder-decoder network. This involved assembling the rs-fMRI data and the corresponding visual stimulus images to create the large-scale triplets used to train the model. At the output end, I implemented the downstream evaluation pipeline to validate the biological relevance of the synthesized activations. I utilized MindEye2 as a downstream probe to reconstruct images from the predicted trial-wise activation volumes.