IIT Mandi
Researcher
Role Overview & Impact
- Designed foundational models for 3D streamlines using PointGPT style architectures with novel tokenization, dual-masking pretraining, and multi-embedding representation learning for WM tract segmentation.
- Built and deployed SlicerTracto, an open-source 3D Slicer extension for brain MRI tractography (reconstruction, tracking, segmentation, visualization); deployed at PGI Chandigarh Hospital.
- Published 5 papers at ICPR'24 (TractoEmbed, TractRL-former), ISBI'25 (TractoGPT), and ISBI'26 (TrackletGPT, TractRLFusion), outperforming Harvard Medical School (TractCloud) and SCIL Canada (FIESTA).
Projects & Initiatives at IIT Mandi
Detailed technical models and systems built during this tenure
Multimodal Text & Sketch-to-3D point cloud generation and interactive editing framework using SDFusion + LLaMA/GPT continuous feedback loops for granular region-specific 3D spatial manipulation. Mentored B.Tech thesis project.
Open-source 3D Slicer desktop extension enabling Neuro-Radiologists to visualize, track, and segment white matter tracts from Diffusion MRI scans in a standalone GUI app with optional on-prem/cloud GPU compute offloading. Deployed at PGI Chandigarh.
TrackletGPT
Accepted at IEEE International Symposium on Biomedical Imaging (ISBI 2026, London, UK): TrackletGPT (Conference Rank: A). Introduces B-Spline streamline sub-segment tokenization for GPT-based 3D point cloud reconstruction and segmentation.
Accepted at 27th ICPR 2024 (Kolkata): Tract-RLformer. Supervised + Reinforcement Learning 2-stage policy refinement network that directly delineates white matter tracts, outperforming SCIL Canada benchmarks.
Preliminary research project establishing spatial 3D streamline patch tokenization that laid the foundational groundwork for TractoGPT, TractoEmbed, and TrackletGPT.
Accepted at 27th ICPR 2024 (Kolkata): TractoEmbed. Modular multi-embedding framework using CNN, dVAE, and PointNet. Outperformed Harvard Medical School's TractCloud benchmark.
Interactive 3D simulation of Chandrayaan-3 Vikram Rover lunar landing and surface traversal built in Unity Engine for mobile VR under Dr. Varun Dutt.