Find Your Daughter

Inspiration

Branchseed Challenge:
Given a 3D CT volume and a binary segmentation mask containing only the parent abdominal aorta, detect every eligible artery that directly branches off the aorta and return each detected branch as an independent, validated daughter instance.

What it does

Find Your Daughter ingests raw CT scans alongside a binary aorta mask to reliably identify, trace, and characterize direct daughter arteries leaving the parent vessel.

For each detected daughter branch, the system computes and extracts:

• Origin point: Where the artery detaches from the aortic wall.
• Seed point & radius: Precise 3D vessel lumen coordinates and estimated vessel caliber.
• Trajectory & orientation vector: Directional cosine vectors of the outbound branch.
• Physical-space mapping: Converts voxel-space indices $(i, j, k)$ into millimeter-space world coordinates $(x, y, z)$ based on scan origin and spacing matrices.

To make the solution accessible and inspectable, the platform features an interactive Streamlit dashboard equipped with slice-by-slice multi-planar overlays and 3D volumetric visualizations to inspect detected branches in real time.

How we built it

• Vascular Analysis & Backend: Developed in Python leveraging SimpleITK, NumPy, SciPy, and scikit-image. The segmentation and detection workflow combines Hounsfield Unit (HU) windowing/intensity thresholding, 3D Medial Axis skeletonization, multi-scale Frangi/vesselness filtering, graph-based topological path tracing, and morphological boundary dilation to locate branch detachment zones.
• Modular Pipeline: We built and benchmarked four detection strategies:
  1. Baseline Model: Fast morphological contact and threshold-based perimeter searching.
  2. Contact Model: Boundary-neighborhood dilation tracking exit surfaces along the aorta wall.
  3. Fusion Model: Combines contact geometry with vesselness eigen-analysis and graph tracing to prune false positives.
  4. Refined Detection Model: Advanced iterative refinement pipeline that enhances seed localization and trajectory estimation by filtering false attachments, optimizing centerline graph tracing, and applying stricter anatomical boundary constraints.
• Synthetic CT Generator: To train, benchmark, and stress-test the pipeline without patient privacy constraints, we engineered a procedural CT data generator. It simulates realistic, contrast-enhanced curved and tapered vascular trees with anatomically constrained branching angles, surrounded by simulated anatomical structures (ribs, spine, lungs, liver, kidneys, and bowel) with Hounsfield-scale noise and blur artifacts.

Challenges we ran into

• Steep Medical & Anatomical Learning Curve: Coming in with minimal clinical imaging experience, our team had to quickly grasp vascular anatomy (e.g., distinguishing celiac trunks, renal arteries, and mesenteric branches from surrounding vascular noise) and radiological conventions.
• Scarcity of Ground-Truth Labeled Data & Synthetic Fidelity: We were provided with very little labeled training data, which severely limited direct supervised validation. To overcome this, we had to build our own procedural synthetic CT generator from scratch. Simulating realistic vascular anatomy was an immense challenge—we had to carefully model smooth 3D vessel tapering, anatomically plausible branching angles, surrounding organ geometry, and realistic Hounsfield Unit noise profiles without introducing non-physiological artifacts that could bias our models.
• False Positive Pruning: High-contrast structures abutting the aorta (such as vertebrae, calcified plaques, or adjacent veins) frequently triggered false boundary detections, requiring us to design strict geometric directionality and continuity checks.

Accomplishments that we're proud of

• Strict Interface Compliance: Fully preserved the required Branchseed evaluator schema and automated validation contracts.
• End-to-End Precision: Successfully implemented physical-space JSON outputs complete with calibrated millimeter radii, seed points, and 3D unit direction vectors.
• Tri-Model Architecture: Shipped four specialized detection models tailored for different speed-versus-recall trade-offs.
• Full-Stack Tooling: Delivered both an interactive Streamlit exploration suite and automated regression testing scripts for synthetic benchmarks.

What we learned

• Building responsive, multi-view medical exploration dashboards using Streamlit.
• Formulating 3D volumetric image processing algorithms, including Hessian matrix vesselness analysis, morphological topology thinning, and graph traversal.
• Handling standard medical imaging formats (NIfTI) and mastering spatial metadata transformations between image voxels and physical patient space.

What's next for Find Your Daughter

• Deep Learning Hybridization: Integrate lightweight 3D U-Net or V-Net segmentation backbones to assist downstream branch tracing in noisy or low-contrast scans.
• Expanded Clinical Evaluation: Benchmark against large-scale clinical registries and public challenges (such as KiTS or MSD datasets).

Built With

• python
• scikit-image
• simpleitk
• streamlit
• vtk