We might have noticed that there are not only texture differences but also color differences due to different stain protocols in the train and test set. So, I wanted to create this post to have an overview of varying normalization techniques for histopathological images.
\nI created this notebook: Stain Normalization techniques: StainGAN, StainNet to illustrate the following techniques:
\nHere is a recap of the data description:
\nHPA samples: stained with 3,3'-diaminobenzidine (DAB) + hematoxylin (H)
\nHuBMAP samples: stained with hematoxylin and eosin (H&E)
\n-> reference from Darien Schettler discussion on Understanding the Different Tissue Stains
There are many histological image normalization methods developed explicitly for H&E staining. However, only a few studies investigated DAB+H staining.
\nThere are 3 groups of conventional stain normalization methods [reference: Complete normalization]
\n1) Histogram matching
\n2) Color transfer
\n3) Spectral matching
\nHowever, the limitation in these conventional methods is that histological information is hardly preserved after normalization.
Here are a few common stain normalization techniques developed by Reinhard, Macenko, and Vahadane. [reference: A review of stain removal]
\nAnd of course, they have deep-learning-based methods primarily applying generative adversarial networks (GANs) as well.
\nSome previous works:
To sum up, we can try out these conventional and deep learning-based normalization techniques as the preprocessing step for this competition. We might want to compare the performances of those techniques with a baseline model to understand which method might work better.
\nLet me know if anything is missing or if I need to make any corrections. 😊
", "messages": [ { "id": "1864156", "postDate": "07/20/2022 19:06:10", "content": "We might have noticed that there are not only texture differences but also color differences due to different stain protocols in the train and test set. So, I wanted to create this post to have an overview of varying normalization techniques for histopathological images.
\nI created this notebook: Stain Normalization techniques: StainGAN, StainNet to illustrate the following techniques:
\nHere is a recap of the data description:
\nHPA samples: stained with 3,3'-diaminobenzidine (DAB) + hematoxylin (H)
\nHuBMAP samples: stained with hematoxylin and eosin (H&E)
\n-> reference from Darien Schettler discussion on Understanding the Different Tissue Stains
There are many histological image normalization methods developed explicitly for H&E staining. However, only a few studies investigated DAB+H staining.
\nThere are 3 groups of conventional stain normalization methods [reference: Complete normalization]
\n1) Histogram matching
\n2) Color transfer
\n3) Spectral matching
\nHowever, the limitation in these conventional methods is that histological information is hardly preserved after normalization.
Here are a few common stain normalization techniques developed by Reinhard, Macenko, and Vahadane. [reference: A review of stain removal]
\nAnd of course, they have deep-learning-based methods primarily applying generative adversarial networks (GANs) as well.
\nSome previous works:
To sum up, we can try out these conventional and deep learning-based normalization techniques as the preprocessing step for this competition. We might want to compare the performances of those techniques with a baseline model to understand which method might work better.
\nLet me know if anything is missing or if I need to make any corrections. 😊
", "rawMarkdown": "We might have noticed that there are not only texture differences but also color differences due to different stain protocols in the train and test set. So, I wanted to create this post to have an overview of varying normalization techniques for histopathological images.\n\nI created this notebook: [Stain Normalization techniques: StainGAN, StainNet](https://www.kaggle.com/code/nghihuynh/stain-normalization-staingan-stainnet) to illustrate the following techniques:\n\n---\n\nHere is a recap of the [data description](https://www.kaggle.com/competitions/hubmap-organ-segmentation/data):\n+ Train set: only contains public HPA data\n+ Public test set: contains private HPA and HuBMAP data\n+ Private test set: contains only HuBMAP data\n\n**HPA** samples: stained with 3,3'-diaminobenzidine (**DAB**) + hematoxylin (**H**)\n**HuBMAP** samples: stained with hematoxylin and eosin (**H&E**)\n-> [reference from Darien Schettler discussion on Understanding the Different Tissue Stains ](https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/333704)\n\nThere are many histological image normalization methods developed explicitly for H&E staining. However, only a few studies investigated DAB+H staining. \n\nThere are 3 groups of **conventional stain normalization methods** [[*reference: Complete normalization*](https://www.ece.ualberta.ca/~xingyu/resources/J_TBME2015.pdf)]\n1) Histogram matching\n2) Color transfer\n3) Spectral matching\nHowever, the limitation in these conventional methods is that histological information is hardly preserved after normalization.\n\nHere are a few common stain normalization techniques developed by Reinhard, Macenko, and Vahadane. [[*reference: A review of stain removal*](https://iopscience.iop.org/article/10.1088/1742-6596/1362/1/012108/pdf)]\n+ **Reinhard**: based on color transfer between a standard image and color varied image using mean and variance of both the images. Then. alter the source image to the target image\n+ **Macenko**: find particular stain vectors for each image based on the colors that are present in the image\n+ **Vahadane:** decompose the image into stain density map that are sparse and non-negative. Then, the stain density maps are combined on the basis of stain color of a pathologist preferred target image. Thus, altering only its color and preserving the structure.\n\nAnd of course, they have **deep-learning-based methods** primarily applying generative adversarial networks (GANs) as well.\nSome previous works:\n+ **StainGAN:** based on CycleGAN to transfer the stain style. Limitation: complex and might have a risk of introducing some artifacts [[*reference: Cycle-consistent adversial networks*](https://arxiv.org/abs/1703.10593)]\n+ **StainNet:** uses StainGAN as the teacher network, to learn the color mapping by distillation learning. [[*reference: StainNet-a fast and robust network*](https://www.frontiersin.org/articles/10.3389/fmed.2021.746307/full)] \n\n---\n\nTo sum up, we can try out these conventional and deep learning-based normalization techniques as the preprocessing step for this competition. We might want to compare the performances of those techniques with a baseline model to understand which method might work better. \n\n---\n\nLet me know if anything is missing or if I need to make any corrections. 😊", "votes": null }, { "id": "1868353", "postDate": "07/23/2022 23:35:37", "content": "Great post!
", "rawMarkdown": "Great post!", "votes": null }, { "id": "1870414", "postDate": "07/25/2022 14:31:11", "content": "Thanks Devastator!
", "rawMarkdown": "Thanks Devastator!", "votes": null }, { "id": "1870689", "postDate": "07/25/2022 18:15:48", "content": "Important note:
\nEven though stain normalization has some proven benefits for training our models, experts in this field recommend that it's not always useful. Since we want to build a model being robust to color, it's better to do color augmentation instead of normalizing them. However, you can still use it as long as it's reasonable.
Nice post.
", "rawMarkdown": "Nice post.", "votes": null } ], "comments": [ { "id": 1868353, "author_name": "thedevastator", "author_url": "", "post_date": "07/23/2022 23:35:37", "content": "Great post!
", "votes": null, "replies": [ { "id": 1870414, "author_name": "nghihuynh", "author_url": "", "post_date": "07/25/2022 14:31:11", "content": "Thanks Devastator!
", "votes": null, "replies": [] } ] }, { "id": 1870689, "author_name": "nghihuynh", "author_url": "", "post_date": "07/25/2022 18:15:48", "content": "Important note:
\nEven though stain normalization has some proven benefits for training our models, experts in this field recommend that it's not always useful. Since we want to build a model being robust to color, it's better to do color augmentation instead of normalizing them. However, you can still use it as long as it's reasonable.
Nice post.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1864156": "We might have noticed that there are not only texture differences but also color differences due to different stain protocols in the train and test set. So, I wanted to create this post to have an overview of varying normalization techniques for histopathological images.\n\nI created this notebook: [Stain Normalization techniques: StainGAN, StainNet](https://www.kaggle.com/code/nghihuynh/stain-normalization-staingan-stainnet) to illustrate the following techniques:\n\n---\n\nHere is a recap of the [data description](https://www.kaggle.com/competitions/hubmap-organ-segmentation/data):\n+ Train set: only contains public HPA data\n+ Public test set: contains private HPA and HuBMAP data\n+ Private test set: contains only HuBMAP data\n\n**HPA** samples: stained with 3,3'-diaminobenzidine (**DAB**) + hematoxylin (**H**)\n**HuBMAP** samples: stained with hematoxylin and eosin (**H&E**)\n-> [reference from Darien Schettler discussion on Understanding the Different Tissue Stains ](https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/333704)\n\nThere are many histological image normalization methods developed explicitly for H&E staining. However, only a few studies investigated DAB+H staining. \n\nThere are 3 groups of **conventional stain normalization methods** [[*reference: Complete normalization*](https://www.ece.ualberta.ca/~xingyu/resources/J_TBME2015.pdf)]\n1) Histogram matching\n2) Color transfer\n3) Spectral matching\nHowever, the limitation in these conventional methods is that histological information is hardly preserved after normalization.\n\nHere are a few common stain normalization techniques developed by Reinhard, Macenko, and Vahadane. [[*reference: A review of stain removal*](https://iopscience.iop.org/article/10.1088/1742-6596/1362/1/012108/pdf)]\n+ **Reinhard**: based on color transfer between a standard image and color varied image using mean and variance of both the images. Then. alter the source image to the target image\n+ **Macenko**: find particular stain vectors for each image based on the colors that are present in the image\n+ **Vahadane:** decompose the image into stain density map that are sparse and non-negative. Then, the stain density maps are combined on the basis of stain color of a pathologist preferred target image. Thus, altering only its color and preserving the structure.\n\nAnd of course, they have **deep-learning-based methods** primarily applying generative adversarial networks (GANs) as well.\nSome previous works:\n+ **StainGAN:** based on CycleGAN to transfer the stain style. Limitation: complex and might have a risk of introducing some artifacts [[*reference: Cycle-consistent adversial networks*](https://arxiv.org/abs/1703.10593)]\n+ **StainNet:** uses StainGAN as the teacher network, to learn the color mapping by distillation learning. [[*reference: StainNet-a fast and robust network*](https://www.frontiersin.org/articles/10.3389/fmed.2021.746307/full)] \n\n---\n\nTo sum up, we can try out these conventional and deep learning-based normalization techniques as the preprocessing step for this competition. We might want to compare the performances of those techniques with a baseline model to understand which method might work better. \n\n---\n\nLet me know if anything is missing or if I need to make any corrections. 😊", "1868353": "Great post!", "1870414": "Thanks Devastator!", "1870689": "**Important note:** \nEven though ***stain normalization*** has some proven benefits for training our models, experts in this field recommend that it's ***not always useful.*** Since we want to build a model being robust to color, it's ***better*** to do ***color augmentation*** instead of normalizing them. However, you can still use it as long as it's reasonable.", "1935423": "Nice post." }, "source": "meta" }