Histopathology-Stain-Color-Normalization

Deep Convolutional Gaussian Mixture Model for Stain-Color Normalization in Histopathological H&E Images. The TensorFlow GPU implementation.

Overview

Stain-color variation degrades the performance of the computer-aided diagnosis (CAD) systems. In the presence of severe color variarion between training set and test set in histopathological images, current CAD systems including deep learning models suffer from such an undesirable effect. Stain-color normalization is known as a remedy.

Methodology

Stain-color normalization model can be defined as a generative models that by applying on input image can create different color copies of input image to somehow the converted image contain specific chromatic distribution. Our proposed method contains two stage: (1) Fitting a Gaussian mixture model (GMM) by considering the shape and apearance of image content structures. To do so the visual representation and modeling of convolutional neyral networks (CNNs) are exploited. (2) transforming the estimated distribution to any arbitary distribution that computed from a secondary (template) image.

Features

Fully unsupervised end-to-end learning algorithm
Best performance in color constancy in normalized images
Lack of any threshold and prior assumption about the image contents
Considering image-content structures for fitting a color distribution to the images.

*The tissue class membership, computed by the standard GMM algorithm (middle) and the DCGMM (right); Clusters include nuclei (red), surrounding tissues (green) and the background(blue).*

How to use the code

Prerequisit

Configurations

Specify the dataset diectory containing image patches crpped in predefined with and height.
Specify the logging directory for storing the model trained parameters.
Set the mode switch "mode" in train or prediction. [train]: training the DCGMM for learning the color distribution among images. [prediction]: Use the learned model for color conversion.