Isotropic 3D Bio-Imaging Reconstruction

Project Overview

In fluorescence and confocal microscopy of biological tissue, axial resolution ($z$) is typically 3 to 5 times coarser than lateral in-plane resolution ($xy$). This asymmetry (anisotropy) introduces severe geometric distortion when reconstructing three-dimensional volumes of cell nuclei, membranes, and biliary canaliculi.

To resolve this limitation, I designed and trained an unpaired generative translation pipeline using CycleGAN in PyTorch, translating low-resolution axial slices ($xz$) into high-fidelity lateral representations ($xy$) without requiring paired ground-truth acquisitions.

---

Technical Challenges & Solutions

• Massive Data Ingestion Pipeline: Engineered a high-throughput data loader to extract and preprocess over 120,000 image patches ($128 \times 128$ resolution) from 12 distinct tissue microscopy datasets stored as compressed .npz tensor archives.
• Cycle Consistency Constraint: Dual cycle-consistency losses ensured morphological fidelity, preventing visual hallucinations that could compromise clinical research integrity.
• Rigorous LaTeX Documentation: Mathematical formulations, loss weighting dynamics, and experimental benchmarks were compiled into technical reports formatted in LaTeX for reproducible research.