Projects

Computer use agents are expensive and unreliable across long context windows due to the vision + language + agentic complexity and the cost and unreliability of VLM agents compounds across long running steps. This project aims to build on top of Microsoft's OmniParser architecture and explore integration of implicit brain signals as a new source of data to augment and validate computer use agents' actions and to make computer use agents more reliable and safe across long multi-step scenarios.

Small midbrain nuclei such as the substantia nigra (SN) and ventral tegmental area (VTA) are difficult to localize reliably using template-based atlases alone, particularly across individuals. Neuromelanin-sensitive MRI (NM-MRI) provides subject-specific contrast that may help improve localization of these dopaminergic structures.

In this project, we aim to develop an open and modular workflow that incorporates NM-MRI as a subject-informed prior to refine SN and VTA region-of-interest masks in native space. These masks could then be used for applications such as diffusion tractography, connectivity analyses, seed-based fMRI, or quality control of midbrain segmentations.

During Brainhack, we will prototype the workflow, implement basic QC and evaluation tools, and make the code openly available. We welcome participants interested in neuroimaging, Python, MRI analysis, or open science.

Diffusion Weighted MRI (dMRI) is a cool technology that has improved our understanding of brain microstructure and disease. In research, there are multiple approaches to model microstructure from dMRI data. For example, Diffusion Kurtosis Imaging (DKI) calculates quantitative metrics that potentially explain the brain's complex microstructural configuration. Here at Western, members of the CFMM have developed improvements for DKI making it more robust to noise while also reducing their acquisition time in the MRI scanner. However, these implementations were done on a closed platform (Matlab). This project aims to disseminate open science practices and research done here at Western, integrating these developments into the open source DIPY ecosystem (Python). We hope to give back powerful tools to the neuroscience community to tackle complex questions with dMRI. \n\nSkills for project: Familiarity with matlab, python, git, dMRI (But not really, everyone is welcome).

Triton is a modular audio utility that standardizes stimuli preparation and signal degradation for speech research. It bridges raw signal math with accessible lab tools through three integrated components: a Python engine for RMS-based SNR mixing and vocoding, a CLI for batch processing entire audio directories, and a Streamlit GUI with drag-and-drop degradation and visualization. Built with Pixi for environment reproducibility, Triton ensures consistent audio processing across machines and operating systems—eliminating a major pain point in auditory science research where stimuli preparation often varies between labs or researchers.

This project develops the Automatic Language Mapping Pipeline (ALMP) a fully automated tool for analyzing functional MRI (fMRI) data used in presurgical language mapping to identifying language-related brain regions . ALMP standardizes preprocessing and first-level analysis using MATLAB and SPM12 while providing a user-friendly graphical interface that guides users through the workflow. The pipeline automatically converts imaging data, performs motion correction, slice-timing correction, coregistration, and model estimation to generate activation maps of language networks.

MatMRI is a GPU-enabled MATLAB package for advanced MRI reconstruction, modeling, and quantitative analysis, supporting both Cartesian and non-Cartesian data. It provides tools for non-Cartesian regridding, iterative SENSE reconstruction, and higher-order signal modeling, including B0 inhomogeneity correction and time-varying phase estimation using spherical harmonics informed by field monitoring.

This project focuses on improving bview, MatMRI’s image viewer and ROI tool. While it already supports image inspection, ROI drawing, and quick SNR calculations on selected ROIs, planned enhancements aim to make bview fMRI-focused. Improvements include interactive voxel time series visualization by clicking on a pixel (with options to normalize, demean, or convert to percent signal change), generation and export of tSNR maps, and the ability to overlay activation maps on reconstructed images. Additionally, we plan to extend bview to inspect processed field monitoring outputs, enabling k-space trajectory visualization and validation against the nominal trajectory.

Traditional neuroimaging analyses focus on visual features, which makes it difficult to portrait continuous clinicopathological changes. To facilitate the synergetic analyses of similar records, we attempt to use constrative learning, a cutting edge self-supervised deep learning framework, to map the individual MRI features into a high-dimensional manifold with reference to the topology correlations in their pathology records. Specifically, it requires a specialized similarity metric incooperating both imaginery and graphical informations. The expected outcome is an interactive web interface tool to visualize the disease's evolutionary trajectory and pinpoint pathogenic brain regions.

Functional MRI is a fascinating approach to study how brain networks develop in early childhood and what factors in the early environment shape child developmental trajectories. In a randomized controlled trial of an intervention for mothers with postpartum depression, we acquired child (age 3-7) functional neuroimaging data in a subset of children. We have noticed some reconstruction artifacts at the top and bottom of the head in some fMRI data that had motion, and these slices need to be realigned for standard motion processing pipelines to work well.

SPIMquant's registration pipeline (affine + deformable via greedy) takes minutes per subject and is prone to failure modes. We replace stages 1–6 with a single 3D U-Net that directly predicts atlas labels from downsampled SPIM images in seconds. Inspired by SynthSeg (Billot et al.), we train on purely synthetic images generated on-the-fly from 52 existing SPIMquant-derived label maps — randomizing intensity, deformation, bias field, noise, and blur so the network generalizes across contrasts and acquisitions without ever seeing a real training image. Skills: Python, PyTorch/MONAI, neuroimaging (NIfTI/nibabel), deep learning, git Repo: github.com/Arshya-Guru/SPIMDash