LuGot16/tubule-segmentation-unet

Seminiferous Tubule Segmentation Model

U-Net with ResNet34 encoder for 3-class semantic segmentation of seminiferous tubules in H&E-stained histology images.

Classes

ID	Class	Color in overlays	Description
0	Background	—	Interstitial tissue
1	Tubule Wall	🟢 Green	Seminiferous epithelium
2	Lumen	🔵 Blue	Hollow center of tubule

Results

Evaluated on 70 validation images:

Metric	Mean ± Std
Mean IoU	0.7498 ± 0.0188
IoU Background	0.9634 ± 0.0192
IoU Tubule Wall	0.4255 ± 0.0313
IoU Lumen	0.8603 ± 0.0300
Dice Background	0.9813 ± 0.0105
Dice Tubule Wall	0.5963 ± 0.0316
Dice Lumen	0.9246 ± 0.0174

Model Details

Architecture	U-Net (segmentation_models_pytorch)
Encoder	ResNet34 (ImageNet pretrained)
Input	RGB, 256×256, ImageNet normalization
Output	3-class segmentation mask
Parameters	~24.4M
Best epoch	42
Checkpoint IoU	0.8677

Training Configuration

• Dataset: LuGot16/tubules-segmentation (281 train / 70 val)
• Loss: 0.5×DiceLoss + 0.5×CrossEntropyLoss (inverse-frequency class weights)
• Optimizer: AdamW (lr=1e-4, weight_decay=1e-4)
• Scheduler: CosineAnnealingLR
• Augmentations: RandomResizedCrop, D4 symmetry (flips + rot90), ElasticTransform, ColorJitter

Quick Start

import torch
import segmentation_models_pytorch as smp
from huggingface_hub import hf_hub_download

# Load
ckpt_path = hf_hub_download("LuGot16/tubule-segmentation-unet", "best_model.pt")
ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)
model = smp.Unet(encoder_name="resnet34", encoder_weights=None, in_channels=3, classes=3)
model.load_state_dict(ckpt["model_state_dict"])
model.eval()

Inference with Area Calculation

from inference import load_model, predict

model, config = load_model("LuGot16/tubule-segmentation-unet")
mask, areas = predict(model, "image.png", config)

print(f"Tubule wall area: {areas['tubule_wall_area_um2']:.1f} μm²")
print(f"Lumen area: {areas['lumen_area_um2']:.1f} μm²")
print(f"Lumen/Tubule ratio: {areas['lumen_to_tubule_ratio']:.3f}")
print(f"Tubule coverage: {areas['tubule_coverage_pct']:.1f}%")

Default scale: 0.32 μm/pixel (calibrated from 25μm scale bar). Adjust with scale_um_per_px parameter.

Files

• best_model.pt — Trained model checkpoint
• inference.py — Inference script with area calculation (μm²)
• extract_masks.py — Tool to extract GT masks from red-contour-annotated images
• train.py — Training script (reproducible)
• evaluate.py — Evaluation script
• visualizations/ — GT vs prediction comparisons on val set

Data Pipeline

1. Original images have manually drawn red contours marking tubule boundaries
2. extract_masks.py uses OpenCV contour hierarchy detection (RETR_CCOMP) to extract 3-class masks
3. Red annotations are removed via inpainting for clean training inputs
4. Model is trained on clean images, predicts on unannotated images