Go to MICCAI 2019 Workshop COMPAY homepageStain Standardization Capsule: A pre-processing module for histopathological image analysis 
Submission by Yushan Zheng • Stain Standardization Capsule: A pre-processing module for histopathological image analysis
Keywords: Stain standardization, Histopathological image analysis, Digital pathology, Capsules network
TL;DR: A novel stain standardization module for histopahtological image normalization is proposed.
Abstract: Color consistency is crucial to developing robust deep learning methods for histopathological image analysis. With the increasing application of digital histopathological images, the deep learning methods are likely developed based on the data from multiple medical centers. This requirement makes it a challenge task to normalize the color variance of histopathological images from different medical centers. In this paper, we proposed a novel color standardization module named stain standardization capsule (SSC) based on the paradigm of capsule network and the corresponding dynamic routing algorithm. The proposed module can learn and generate uniform stain separation outputs for histopathological images in various color appearance without the reference to manually selected template images. The SSC module is light and can be trained end-to-end with the application-driven CNN model. The proposed method was validated on two public datasets and compared with the state-of-the-art methods. The experimental results have demonstrated that the SSC module is effective in color normalization for histopathological images and achieves the best performance in the compared methods.
Reply Type:
Author:
Visible To:
Hidden From:
4 Replies
[–][+]
Stain Standardization Capsule: A pre-processing module for histopathological image analysis
Official Review of Paper18 by AnonReviewer1 • Stain Standardization Capsule: A pre-processing module for histopathological image analysis
Review: Clarity: The authors can provide more clarity on how the training is done.
Summary: Color standardization module, based on capsule network and dynamic routing algorithm is proposed in the context of follow-up CNN for tumor cell-classification application. The combined network is trained together and evaluated, using two public datasets of Camelyon-16 and ACDC-LungHP, and compared with alternative methods, including the one without any color standardization.
Results: SSC model is shown to be relatively slightly better in regards to classification metrics with fewer number of parameters and not requiring additional manual input. (From the reported results, I notice that the results from different methods are quite similar and the method with any color standardization is not all that bad - which leads me to guess that the classification CNN network has the implicit capacity to model and accommodate the stain color variability. So, that raises the broader question - is color normalization required for CNN based histopathological image analysis).
Originality: SSC module is an interesting and attractive approach as: 1) it makes use of the stain formation process, by building the network on OD input and the number of network parameters closely mimics the number of physical reference dye vectors used in the staining process 2) Addresses the relatively common issue on how to handle stained slides resultant from different staining protocols.
Pros/Cons:
The paper is incomplete. What would be interesting follow up questions to be pursued are:
A) understand the differences in image reconstruction appearance and specific trends/patterns in the classification errors from different methods.
B) How consistent and effective is this method (or other methods) for other HIA applications :
1) like how does it impact the image appearance in the WS image ( an important question for manual reading)
2) in algorithmic detection and classification of other micro and macro structures (such as lymphocytes, blood vessels, macrophages etc, glandular structures and muscular structures) for various tissue type slides.
C) The provided quantitative results and image patches are not adequate enough to conclusively show the improvements of the proposed method over others, in regards to the important challenges in automated interpretation of slides with staining variations mentioned in the introduction section. The authors need to provide specific esults (may be online) for couple of whole slide images and different tissue types to show that the color normalization method is effective.
Significance:
The work can generate some interesting discussion on A) the need for color standardization in the context of CNN-based classification methods. B) Methods and impact of color normalization on downstream automated H&E analysis for various tissue types and manual reads of digitized slides.
Rating: 7: Good paper, accept
Confidence: 3: The reviewer is fairly confident that the evaluation is correct
[–][+]
Review
Official Review of Paper18 by AnonReviewer4 • Review
Review: This paper is focusing on an important concept in medical imaging: image standardization/normalization which is important for suppressing difference caused by data preparation or imaging procedure and let the deep learning pipeline to focus on application purposes.
Besides evaluating classification benefits, could authors directly evaluate the quality for standardization?
In the step of SSC, is it possible to skip the step of generating Ui ?
For people who are not familiar with DR , please elaborate this.
It is not clear the motivation of bi in Algorithm 1, please explain.
Rating: 6: Marginally above acceptance threshold
Confidence: 3: The reviewer is fairly confident that the evaluation is correct
[–][+]
Interesting methodology.
Official Review of Paper18 by AnonReviewer2 • Interesting methodology.
Review: Quality: Well written and well-structured
Clarity: Generally clear. The experimental setting is not clear: what is the classification task for each dataset? The number of samples? How do the training and validation perform? It's not clear that the validation between different algorithm is fair. If the proposed method been trained end-to-end for the classification task, have other been trained end-to-end?
Originality: The application of capsule net to stain standardization is novel and the use of sparsity constrains in the dynamic routing is very original.
Significant of this work: High because it discards the need for the target colour distribution.
Pros:
- A very attractive approach to stain standardization: 1) less no. parameters, 2) no need of a template, and 3) can be trained end-to-end to a specific task.
Cons:
- The details for the experimental setting need to more clarity. See above.
Rating: 7: Good paper, accept
Confidence: 4: The reviewer is confident but not absolutely certain that the evaluation is correct
Paper Decision
Decision by Program Chairs • Paper Decision