Seminiferous Tubule Segmentation — Coarse-to-Fine Model (v20b)

Model Description

A coarse-to-fine dual-branch architecture for 3-class semantic segmentation of mouse seminiferous tubules in H&E-stained histology images.

Classes: Background (0) | Seminiferous Epithelium (1) | Lumen (2)

Architecture

Branch	Encoder	Input	Purpose
Coarse	ResNet34	Full image 512×512	Identifies the main tubule region
Fine	EfficientNet-B4	Native patch 256×256 + coarse mask	Precise border delineation

The coarse branch provides a soft probability mask that guides the fine branch, ensuring predictions stay within the tubule region. A border detection auxiliary head improves boundary precision.

Total parameters: 47.3M

Training

Loss: Dice + Cross-Entropy (nnU-Net formulation)
Optimizer: AdamW with differential learning rates (GLNet principle)
Schedule: CosineAnnealingWarmRestarts
Augmentation: Kornia GPU (elastic deform, color jitter, Gaussian noise)
Coarse caching: Re-computed every 5 epochs for speed
Dataset: ~3900 patches from 431 images (85/15 train/val split)
Trained from scratch (ImageNet pretrained encoders only)

Usage

import torch
from model import CoarseToFineModel

model = CoarseToFineModel()
ckpt = torch.load('best_model_v20b.pt', map_location='cpu')
model.load_state_dict(ckpt['model_state_dict'])
model.eval()

# See inference.py for full sliding-window inference pipeline

Inference

The model uses a sliding window approach:

Coarse branch processes the full image once → tubule probability map
256×256 patches slide across the image at stride=128
Each patch receives the coarse probability as a 4th input channel
Gaussian-weighted blending of overlapping predictions
Post-processing: largest connected component for epithelium

Scale: 0.32 μm/pixel → areas reported in μm²

Results

Metric	Value
Validation mIoU	TBD (training in progress)
Epithelium IoU	TBD
Lumen IoU	TBD
Test images OK	TBD/51

Hardware

Trained on NVIDIA GTX 1660 Ti (6GB VRAM), AMD Ryzen 5 5600X, 16GB RAM.

Files

├── best_model_v20b.pt       # Model checkpoint
├── src/
│   ├── model.py             # Model architecture
│   ├── dataset.py           # Data loading + patch extraction
│   ├── losses.py            # Dice + CE loss functions
│   ├── augmentation.py      # Kornia GPU augmentation
│   └── utils.py             # Mask extraction utilities
├── train.py                 # Training script
├── inference.py             # Sliding window inference
└── evaluate.py              # Test set evaluation

Citation

@software{seminiferous_segmentation_2025,
  title={Coarse-to-Fine Segmentation of Seminiferous Tubules},
  author={Lucila Gotfryd},
  year={2025},
  url={https://huggingface.co/LuGot16/seminiferous-tubule-segmentation}
}

License

Apache-2.0

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

Try ML Intern: https://smolagents-ml-intern.hf.space
Source code: https://github.com/huggingface/ml-intern

Downloads last month: -; Downloads are not tracked for this model. How to track

LuGot16
/

seminiferous-tubule-segmentation