LuGot16/tubules-segmentation

Seminiferous Tubule Segmentation Dataset

Processed dataset for 3-class semantic segmentation of seminiferous tubules in H&E-stained histology images.

Classes

• 0 - Background: Interstitial tissue and non-tubule regions
• 1 - Tubule wall: Seminiferous tubule wall (epithelium between outer and inner contour)
• 2 - Lumen: Hollow center of the tubule

Dataset Structure

train/
  images/  # Clean tissue images (red annotations inpainted)
  masks/   # Segmentation masks (0=bg, 1=tubule, 2=lumen)
val/
  images/
  masks/
metadata.json

Statistics

• Train: 281 images
• Val: 70 images
• Original source: LuGot16/tubules

Mask Extraction Method

Masks were extracted from manually drawn red contour annotations using:

1. Red pixel detection (R>150, G<100, B<100, R-G>80)
2. Morphological dilation to connect broken lines
3. RETR_CCOMP contour hierarchy detection (outer=tubule boundary, inner=lumen boundary)
4. Contour filling with parent-child relationship mapping