Arpan Chakraborty

Project 3 • Channel: proj-unpaired-diffusion • Unpaired Med-DDPM on BraTS 2021

• Joined Xu Lab’s proj-unpaired-diffusion channel (added by Dr. Xiangrui Zeng) to work with project leaders including Yashwanth Devavarapu and Rakshitha on unpaired MRI-to-MRI generation.
• Contributing to the lab’s proposed VCM v2 modifications: training with limited paired data, higher-resolution volumes, and anatomical / segmentation consistency for MRI-to-MRI synthesis.
• Extended Med-DDPM from a paired to an unpaired setting with a two-stage framework: (1) Coarse Prior Generator — 4-layer UNet with L1 + gradient loss; (2) Conditional Diffusion Refinement — 8-step DDIM.
• Evaluated on BraTS 2021 (150 patients, 2250 slices). Finding: diffusion refinement degrades unpaired performance, consistent with YODA (IEEE TMI 2026).
• Literature completed for SynDiff, CycleDiffusion, and YODA. Next: SynDiff / CycleDiffusion implementations, FID, method comparison table, and downstream nnU-Net Dice / HD95.

Uncertainty Quantification for Medical Image Classification

• Selected for research training in Bayesian Deep Learning, Explainability (XAI), and Out-of-Distribution (OOD) Detection.
• Developing baseline deep learning pipelines with transfer learning for medical scan classification, integrating Bayesian uncertainty estimation techniques.
• Maintaining code and experiment logs in dedicated repository: github.com/ArpanC6/BayesXMed-JU-Research.

Numerical Instability Analysis across Automatic Differentiation (JAX, PyTorch)

• Selected for competitive summer research project with Cornell's SEAL Lab evaluating numerical errors in automated differentiation backends.
• Identified and analyzed Floating-Point Density Bias and Crossover Point Instability in Herbie rewriting algorithms causing Derivative Damage Amplification.
• Extended JAX benchmarks in GradBench (evaluating ba, det, gmm, and nested saddle optimization).
• Maintained mathematical derivations and empirical logs: github.com/ArpanC6/ad-numerical.

Bayesian Inference & SDE Parameter Estimation

• Collaborating on pseudo-likelihood vs optimization-based estimation for stochastic differential equations.
• Contributed stochastic Lotka-Volterra benchmarks to SciMLBenchmarks.jl, patching sampling import and timing issues.

OAuth 2.0 Client Credentials Flow (RFC 6749) for Jenkins SMTP

• Prototyped OAuth 2.0 authentication for Jenkins SMTP to resolve Microsoft Exchange Online authentication issues.
• Built standalone WireMock POC with 14 unit tests, achieving 0.0163ms cache hit latency (300x faster than target).
• 10 merged PRs across the jenkinsci organization and 30+ community code reviews.

Representing IIT Delhi Entrepreneurship Initiatives

• Promoted technical innovation and hackathons, engaging 500+ students across campus ecosystems.