Nucleui segmentation-tensorflow-slim

This source is tensorflow version of [1].
We used tensorflow-slim and implemented on ubuntu 16.04, python3.5, and tensorflow1.13.0rc1.

NOTE

1. To extract nuclei and non-nuclei patch, we used original matlab patch extraction code[3]
   because converted python code[1] decreased segmentation accuracy significantly.
   However, we are still struggling to correct step1_patch_extraction.py python code.

2. Andrew[3] used the modified Alexnet but we use cifarnet included in tensorflow slim instead.
   For this reason, our result of segmentation(middle) are not so good as caffe version(right) like below.

Requirments

python3.5
tensorflow

Current version ran on CPU. Install tensorflow-gpu version and chagne below code if we want to run on GPU.

   cd DEEP_TUTORIAL_ROOT
   gedit step4_train_image_classifier.py

   ------------------------------------------------------
   tf.app.flags.DEFINE_boolean('clone_on_cpu', True,
                            'Use CPUs to deploy clones.')
   change to 

   tf.app.flags.DEFINE_boolean('clone_on_cpu', False,
                            'Use CPUs to deploy clones.')
   -----------------------------------------------------

Overview

Quick start

User can run test step(step5 and step6) with a little modfication in below script using a pre-trained model.
For instance, change CHECKPOINT_DIR variable to the directory where a pre-trained or user-generated model exists.

   cd DEEP_TUTORIAL_ROOT
   ./train_test_nuclei.sh

Data

Download dataset(train and valiation tfRecord) URL_PASSWORD: 1234

   cd DEEP_TUTORIAL_ROOT/data/1-nuclei/images
   mv DOWNLOAD_DIR/nuclei* ./

For step6 segmentation, original image is on here

Training/Evalution

   cd DEEP_TUTORIAL_ROOT/
   step1_patch_extraction.py (not recommended!. use original patch extraction matlab code)
   step2_cross_validation_creation.py 
   step3_generate_datasets 
   step4_train_image_classifier.py 
   step5_eval_image_classifier.py 

Segmentation

   cd DEEP_TUTORIAL_ROOT/
   step6_segment_test_images.py 

Output generation time

It taks to segment an image very long time (almost ~65 minutes / orginal caffe ~75 minutes on one 1080ti GPU)
Fortunately, Andrew[3] reduced processing time considerably in [4].

Acknowledgements

We would like to thank the authors of DLtutorialCode[2], which we use in this work.

References

[1]python version of [3]
[2]https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/slim
[3]original source
[4]http://www.andrewjanowczyk.com/efficient-pixel-wise-deep-learning-on-large-images/