junyuchen245 / unsuprevised_seg_via_cnn Goto Github PK

View Code? Open in Web Editor NEW

68.0 5.0 29.0 2.47 MB

An unsupervised (or self-supervised) loss function for binary image segmentation.

License: MIT License

Python 100.00%

image-segmentation-tensorflow medical-image-segmentation keras-tensorflow unsupervised-learning

unsuprevised_seg_via_cnn's Introduction

Unsupervised Segmentation Using Convolutional Neural Network

This is a Tensorflow/Keras implementation of my paper:

Chen, Junyu, et al. "Medical Image Segmentation via Unsupervised Convolutional Neural Network. " Medical Imaging with Deep Learning (MIDL), 2020.

👋 Note that this method does binary segmentation. Please check out our new approach 👉 (FCM loss) for unsupervised and semi-supervised loss functions for multi-class segmentation (PyTorch and TensorFlow).

Network Architecture:

Evaluation and Results:

We evaluated four settings of the proposed algorithm on the task of bone segmentation in bone SPECT images:

Mode 1: Unsupervised (self-supervised) training with L_ACWE.
Mode 2: Mode 1 + fine-tuning using L_label with 10 ground truth (GT) labels.
Mode 3: Mode 1 + fine-tuning using L_label with 80 GT labels.
Mode 4: Training with L_ACWE + L_label.

The quantitative results can be found in the paper, and here are some qualitative results:

Comparing to traditional ACWE:

If you find this code is useful in your research, please consider to cite:

@inproceedings{
chen2020medical,
title={Medical Image Segmentation via Unsupervised Convolutional Neural Network},
author={Junyu Chen and Eric C. Frey},
booktitle={Medical Imaging with Deep Learning},
year={2020},
url={https://openreview.net/forum?id=XrbnSCv4LU}
}

About Myself

unsuprevised_seg_via_cnn's People

Contributors

Stargazers

Watchers

unsuprevised_seg_via_cnn's Issues

Pytorch

你好，是否有Pytorch的训练代码呢？

Get data set

Thank you very much for opening the source code. Can you provide a little original data set?

Is it possible to extend the model on plain xray of bones

I was wondering since you mentioned 2d slices of the data if this model can be used on 2d images of plain xray of say bones of hands & feet to segment & find regions of interest. Thanks & lovely work

Compare with Chan-Vese model

Dear @junyuchen245 ,

Thanks for sharing the great work.

How about comparing with the naive Chan-Vese Model in the experiments?

The idea of this paper is to minimize the energy functonal of CV model by CNN rather than tradition gradient, am I right?

This work you may be interested in:
Mumford–Shah Loss Functional for Image Segmentation With Deep Learning

Best regards,
Jun

Satellite image semantic segmentation

Is this paper or implementation applicable to Satellite image semantic segmentation?

Could you make the dataset public available ?

Thanks for making the code open-sourced.

May I ask if you can make your training data public available ?

Thank you

Run `Mode4`

Thanks for your great work.

I want to run Mode4 described in your paper.
I ever tried to run other modes successfully.
Could you help me with how to run the Mode4 step by step?

Thanks in advance.

How to train this model on customized data (images)

Hi,

Great work there with some impressive results!

I would like to use your model for my experiments with 3D thermographic images. Could you kindly guide me on how to run the model after cloning this repo?

I understand I have to run train.py as other files are already imported within it.

Regarding the arrangement of my images, how should I go about this? May I know what images and how many of them are in each of these 4 directories? Can I train with just 1 image and test with another?
imgDir = '/netscratch/jchen/SPECTSeg_data/train/'
gtDir = '/netscratch/jchen/SPECTSeg_data/train/'
imgTestDir = '/netscratch/jchen/SPECTSeg_data/test/'
gtTestDir = '/netscratch/jchen/SPECTSeg_data/test/'

Looking forward to hearing from you.

Thanks in advance!