Note: This is the experiment reproduction repository for our JMLR paper.
For the RankSEG software package (installation & usage), please visit:
https://github.com/rankseg/rankseg

RankSEG is a Python module designed for segmentation tasks, aiming to maximize Dice or IoU metrics based on estimated probabilities.

• GitHub repo: https://github.com/statmlben/rankseg
• Slides: https://slides.com/statmlben/rankseg
• Paper: JMLR-v24-22-0712
• Poster: ICML2024

Most segmentation methods traditionally rely on IoU and Dice as evaluation metrics. During inference and prediction, these methods typically use a threshold of 0.5 or apply argmax to the estimated probabilities to generate segmentation predictions. However, this approach does not directly optimize the IoU or Dice metrics.

Our method, RankDice, directly optimizes IoU and Dice metrics.

• Nearly ensures improved Dice and IoU performance!
• Seamlessly integrates with any pretrained segmentation neural network (no need to retrain the models).
• A well-developed Python function rank_dice is available for use.

• Example in VOC 2012 dataset
• Simulated examples

• Threshold, Argmax and rankDice are performed based on the same network (in Model column) trained by the same loss (in Loss column).
• Averaged mDice and mIoU metrics based on state-of-the-art models/losses on Fine-annotated CityScapes val set. '/' indicates not applicable since the proposed RankDice/mRankDice requires a strictly proper loss. The best performance in each model/loss is bold-faced.
• All trained neural networks and their config.json with different network and loss are saved in this link (12G folder: network/loss/.../*.pth + config.json)

• Threshold, Argmax and rankDice are performed based on the same network (in Model column) trained by the same loss (in Loss column).
• Averaged mDice and mIoU based on state-of-the-art models/losses on PASCAL VOC 2012 val set. '---' indicates that either the performance is significantly worse or the training is unstable, and '/' indicates not applicable since the proposed RankDice/mRankDice requires a strictly proper loss. The best performance in each model-loss pair is bold-faced.
• All trained neural networks with different network and loss are saved in this link (22G folder: network/loss/.../*.pth)

• Threshold, Argmax and rankDice are performed based on the same network (in Model column) trained by the same loss (in Loss column).
• Threshold and Argmax are exactly the same in binary segmentation.
• Averaged mDice and mIoU based on state-of-the-art models/losses on Kvasir-SEG dataset set. '---' indicates that either the performance is significantly worse or the training is unstable, and '/' indicates not applicable since the proposed RankDice/mRankDice requires a strictly proper loss. The best performance in each model-loss pair is bold-faced.

• All empirical results on different losses and models can be found here

If you want to replicate the experiments in our papers, please check the folder ./pytorch-segmentation-rankseg and its README file Pytorch-segmentation-rankseg

If you like RankSEG please star the repository and cite the following paper:

If you find this repository helpful, please star our repo .

Thank you so much for the support from our stargazers.