Topo-Consistency

This repository contains the implementation for our work "Semi-supervised Segmentation of Histopathology Images with Noise-Aware Topological Consistency", accepted by ECCV 2024.

Environment

Training and evaluation environment: Python 3.9.12, PyTorch 2.0.1, CUDA 11.7. We use CubicalRipser (cripser) to extract topological features. Run the following command to install required packages.

pip install -r requirements.txt

Inituition of Decomposition & Matching

Overall Framework

Usage

This loss can be incorporated into any teacher-student framework or its variants.

Sample usage: note that we calculate the topological consistency loss between the foreground of the likelihood map.

def calculate_topo_loss(likelihood, target):
  batch_size = likelihood.shape[0]

  topo_loss = 0.0

  for i in range(batch_size):
      lh = likelihood[i]
      #print(lh.shape)
      gt = target[i]

      topo_loss += getTopoLoss(lh, gt)

  topo_loss /= batch_size
  return topo_loss

topo_loss_weight = 0.002
stu_likelihood = torch.softmax(model(unlabeled_data))[:,1,:,:]
tea_likelihood = torch.softmax(teacher_model(unlabeled_data))[:,1,:,:]

topo_consistency_loss = calculate_topo_loss(stu_likelihood, tea_likelihood)
topo_consistency_loss = topo_loss_weight*topo_consistency_loss

Qualitative Results

Citation

If you found this work useful, please consider citing the following articles

@inproceedings{xu2023toposemiseg,
  title={TopoSemiSeg: Enforcing Topological Consistency for Semi-Supervised Segmentation of Histopathology Images},
  author={Xu, Meilong and Hu, Xiaoling and Gupta, Saumya and Abousamra, Shahira and Chen, Chao},
  booktitle={ECCV},
  year={2024}
}

@inproceedings{hu2019topology,
  title={Topology-preserving deep image segmentation},
  author={Hu, Xiaoling and Li, Fuxin and Samaras, Dimitris and Chen, Chao},
  booktitle={NeurIPS},
  volume={32},
  year={2019}
}