Evaluation of the Stain Separation Layer for Her2 Scoring
Evaluating the impact of a stain separation layer for the Her2 Scoring dataset.
Created on May 6|Last edited on July 2
Comment
Introduction
For these section, We evaluate the impact of the color decomposition layer by comparing the HED, HDX, and RGB color spaces as input images of common neural network architecture in a 5-fold cross-validation setup.
Color decomposition, also known as color deconvolution consists of the separation of features by their colors, and the extraction of semantic information relevant to HER2 Scoring task. Due to HER2 Scoring task is correlated with the brown dyed of the cell membranae, a HED color separation may facilitate the recognition of the tissue score.
We propose a layer to incorporate the HED and HDX color decomposition into the network architecture. The proposed layer is a non-trainable convolution layer that computes the color information of each pixel in a new color space. This is achieve by a convolution with a predefined kernel for HED and HDX color spaces.
Patch Classification Model
For HER2 Scoring we use the approach of classify smaller patches obtained from the WSI. For this task, we modified the MC-Dropout model used for the binary classification. In this model, the target consist of a set of four unit, for the 0, 1+, 2+ and 3+ classes respectivly. The second modification is the Encoder submodel, which has an extra layer for the color decomposition. We call this submodel the Stain Separation Encoder.
For the first layer of the Stain Separation Encoder, we add our new preprocessing layer for color decomposition, the Stain Separation Layer, theses patch images of dyed tissue are converted from the RGB color space into the Haematoxylin-Eosin-DAB (HED) color space.
A Her2 Scoring model with MC-Dropout implementation. We include a Stain separation encoder model, in this way we can include the context knowledge into the architecture.
The Stain Separation layer is a non-trainable convolution layer that applies a predefined kernel [1] to RGB image (3 channels) and obtaining an image in the normalized (0-1) HED color spaces. As an alternative, the Eosin channel can be ignored due the information aported may be irrelevant for the HER2 Scoring, the HDX color space can be obtained by replacing the Eosin by a linear combination of Hemaxylin and DAB channels.
A Stain Separation Layer with HED decomposition. The deconvolution can be included as a layer by applying a 1D convolution with a predefined kernel.
Evaluation Results
For this experiment the hyperparameters evaluation consists of the combination of two variables; the former is the Stain Separation Layer color space. We evaluate the impact of applying the color decomposition with three configurations: Hematoxylin, Eosin, and DAB (HED) color separation; Hematoxylin and DAB (HDX) color separation; and no color separation, which we refer to as the RGB color separation.
The latter hyperparameter to evaluate is the classifier configuration. The classifier model is defined by a stack of dense layers with MC-Dropout. For simplicity, we use the notation CXX to represent the number of dense layers, where each X is a dense layer with units followed by an MC-dropout layer. Thus, the configuration C876 represents a dense neural network with 256, 128, and 64 units layers.
The number of possible configurations is enormous. To reduce the exploration space, in this experiment, we will compare only two configurations: C777 and C876. This decision may affect the performance of our models since these configurations haven't been optimized for this task, nonetheless, we can still make the comparison for the color separation model.
The experiment setup is shown in the table below, which includes the color separation and classifier configurations. To compare the generalization of these models we will evaluate each model with a 5-fold cross-validation strategy. Each cross-validation runs a single training instance using a simple pre-specified dataset split. The results are shown in the following figures.
Stain Separation
29
29
The figure shows the average classification metrics for each model configuration in the 5 folds. As expected, models with Stain Separation Layers (HED and HDX) perform well against no applying color deconvolution (RGB). Moreover, HDX models achieve higher classification performance.
In contrast, the results on the classifier configuration don't differ greatly between the two (C77 and C876) in the aggregated view. The following figure shows the comparison of each model configuration.
C876
15
C777
14
15
14
The cross-validation results show that the model HDX-C876 outperforms the other models in the overall performance, but not by a large margin. However, we will use the HDX-C876 model configuration for the following experiments.
Add a comment