Connected-Components Centroid Extraction

Overview

After segmenting a 3D volume into a binary or multi-class mask, convert pixel-level predictions into discrete object detections. Connected-component labeling groups contiguous voxels, then each component's centroid gives the object's 3D coordinate. Essential for particle picking, cell detection, and nodule localization.

Quick Start

import cc3d
import numpy as np

def extract_centroids(segmentation, threshold=0.5):
    binary = (segmentation > threshold).astype(np.uint8)
    labels, n = cc3d.connected_components(binary, return_N=True)
    centroids = []
    for label_id in range(1, n + 1):
        coords = np.argwhere(labels == label_id)
        centroids.append(coords.mean(axis=0))
    return np.array(centroids)  # shape: (N, 3) — z, y, x

Workflow

1. Threshold the probability map to binary mask
2. Run 3D connected-component labeling (cc3d or scipy.ndimage.label)
3. Filter components by volume (discard tiny noise or oversized blobs)
4. Compute centroid of each component via np.argwhere().mean()
5. Output list of (z, y, x) coordinates

Key Decisions

• Connectivity: 26-connected (face+edge+corner) catches thin structures; 6-connected (face-only) is stricter
• Min volume filter: Remove components below N voxels to suppress false positives
• cc3d vs scipy: cc3d is 2-10x faster on large volumes
• Multi-class: Run per-class or use cc3d.connected_components with connectivity=26

References

• Baseline UNet train + submit
• 3d-unet using 2d image encoder