najnaj / deeplearning-digital-pathology

Full package for applying deep learning to virtual slides.

Deeplearning-digital-pathology

This repository contains utilities for virtual slides and images classification and semantic segmentation with Keras and Caffe and an extension class of ImageDataGenerator of Keras to generate batches of images with data augmentation for segmentation. Demo code is provided for reference.

Requirement

Python 2.7
OpenCV 3.4
Numpy 1.14
Tensorflow 1.7
Keras 2.1
OpenSlide 1.1
Caffe 0.15 (optional)

Getting started

slide_demo.py It shows an example of using the Keras or Caffe model to segment a whole virtual slide or classify slide in grid and saving the results.

Segmentation:

Classification:

image_demo.py It shows an example of using the Keras or Caffe model to segment or classify images from a folder and saving the results.

train_segmentation_demo.py It shows an example of training a segmentation model from scratch.

Contents

See comments or use help for detailed function usage.

ImageDataGeneratorEXT.py

An extension of ImageDataGenerator of Keras for semgentation images iteration , with similiar api of flow_from_directory.

The structure of the image directory would be like:

"directory"
├── "image_subfolder"
│   ├── image1.png
│   ├── image2.png
│   └── ...
└── "mask_subfolder"
    ├── image1.png
    ├── image2.png
    └── ...

slide_utils.py

Virtual slide class (Slideobject) for retrieving image batches and results reconstruction.

• retrieve_tiles_to_queue_thread Retrieve image batches and put into a queue waiting for analysis.

• reconstruct_segmentation_queue_to_level Reconstruct segmentation results on top of selected layer image.

• reconstruct_classification_queue_to_level Reconstruct classification results on top of selected layer image.

• preprocess_img Rewrite to include customized image preprocessing.

• gray_code Rewrite based on given code to change color for each category for result_mask.

• color_code Rewrite based on given code to change color for each category for result_rgb.

image_utils.py

Image class (Imageobject) for retrieving image batches and results reconstruction.

• retrieve_tiles_to_queue_thread Retrieve image batches and put into a queue waiting for analysis.

• reconstruct_segmentation_queue_to_file Reconstruct segmentation results on top of original image and save.

• reconstruct_classification_queue_to_file Reconstruct classification results on top of original image and save.

• preprocess_img Rewrite to include customized image pre-processing.

• gray_code Rewrite based on given code to change color for each category for result_mask.

• color_code Rewrite based on given code to change color for each category for result_rgb.

caffe_utils_queue.py

Caffe class (Caffeobject) for forwarding images batch to neural network.

• forward_from_queue_to_queue Forward batch in data queue to neural network and put results in result queue.

tf_utils_queue.py

Tensorflow (Keras) class (TFobject) for forwarding images batch to neural network.

• forward_from_queue_to_queue Forward batch in data queue to neural network and put results in result queue.

Related publications

• Semantic Segmentation of Colon Glands in Inflammatory Bowel Disease Biopsies. Z Ma, Z Swiderska-Chadaj, N Ing, H Salemi, D McGovern, B Knudsen, A Gertych. Information Technologies in Biomedicine, Kamień Śląski, Poland, June 2018.