TractoAware

Preliminary Spatial 3D Point Cloud Streamline Tokenization

Organization: IIT Mandi

Key Result & Impact Benchmark

Validated 3D point cloud patch representations for brain fiber tract spatial modeling.

Technical Overview

TractoAware is the foundational preliminary research project conducted at IIT Mandi's MANAS Lab that initiated the lab's line of 3D Spatial AI publications. The project explored how spatial point cloud representations could be extracted from 3D Diffusion MRI streamline sequences. By grouping neighboring streamlines into spatial 3D point cloud patches, TractoAware demonstrated that self-supervised masked pretraining could capture microstructural brain connectivity patterns, directly inspiring the subsequent TractoEmbed, TractoGPT, and TrackletGPT architectures.

Key Technical Highlights

Established initial proof-of-concept for converting dMRI fiber streamlines into spatial 3D point cloud patches.
Proved viability of transformer pretraining over non-Euclidean neuroimaging spatial structures.
Served as the direct foundational predecessor to 5 peer-reviewed conference publications.

Technologies & Frameworks

3D Spatial AI

Point Cloud Patching

DIPY

PyTorch

Python