SenNet + HOA

U-Net-based 3D vessel segmentation for the SenNet + HOA Kaggle competition.

This project contains my solution for the “SenNet + HOA – Hacking the Human Vasculature in 3D” Kaggle challenge, which focuses on segmenting blood vessels in large 3D TIFF scans of human kidneys.

🔍 Summary

Given massive 3D volumes, the goal is to produce voxel-wise vessel masks that can be used to study vascular morphology and support the Human Reference Atlas effort. The main challenges are:

  • 40GB+ of volumetric data per competition bundle.
  • Highly imbalanced vessel vs. background classes.
  • Tight GPU memory constraints for 3D convolutions.

🧠 Method

  • A 3D U-Net architecture for volumetric segmentation.
  • Training and inference pipelines tailored to:
    • Patch-based processing of large volumes.
    • Careful memory management on commodity GPUs.
  • Post-processing that converts predicted masks to run-length encoded (RLE) strings to match Kaggle’s submission format.

📂 GitHub repository