BS-Loss

This repository is an official PyTorch implementation of the paper "Boundary-Sensitive Loss Function With Location Constraint for Hard Region Segmentation" [paper] (https://ieeexplore.ieee.org/document/9950613) from IEEE Journal of Biomedical and Health Informatics (JBHI) 2022.

Requirements (PyTorch)

Core implementation (to integrate the bs loss into your own code):

• python 3.8
• pytorch 1.8.1
• opencv-python 4.6.0.66
• numpy

To reproduce our experiments:

• python 3.8
• pytorch 1.8.1
• copy
• numpy
• matplotlib
• scikit-image
• random
• glob
• tqdm

Usage

The implementation of BS loss has three key functions:

• the BS loss itself (boundary_sensitive_loss in BS_loss.py#L67;
• the dilation and erosion operations (get_boundary in BS_loss.py#L40);
• the Location Constraint (location_constraint in BS_loss.py#L25);

If you want to use BS loss as the loss function only, you can directly use the BSLoss in BS_loss.py#L6). There is an optional argument in the BSLoss, but we suggest to use the default setting of alpha=0.8. If you want to use BS loss with Location Constraint, you can directly use the BSL_LC in BS_loss.py#L15).

It should be noted that our method is designed for binary segmentation tasks, and you need to ensure that the shape of your network input and output is [N,1,H,W] or [N,H,W].

Cite BS-loss

---

If you used BS loss in your research projects, please remember to cite our reference paper published at the IEEE Journal of Biomedical and Health Informatics (JBHI) 2022. This will help us make BS loss known in the machine learning community, ultimately making a better tool for everyone:

@ARTICLE{9950613,  
author={Du, Jie and Guan, Kai and Liu, Peng and Li, Yuanman and Wang, Tianfu},  journal={IEEE Journal of Biomedical and Health Informatics},   
title={Boundary-Sensitive Loss Function with Location Constraint for Hard Region Segmentation},   
year={2022},  
volume={},  
number={},  
pages={1-12},  
doi={10.1109/JBHI.2022.3222390}}