Slides to Mask using CVAT

[Jimut123]

Probably create a new repository for converting slide images to its corresponding multi-class segmentation mask in using CVAT. We can use our own segmentation model via robust data augmentation methods. We can even crop and cut the dataset with the corresponding masks to generate more and more data for segmentation.

TODO

• Robust Multi-Class Segmentation model
• Convert it to GAN, Probably SegGAN++
• GAN Mask R-CNN
• Cell Image Seg
• Domain adaptation dataset mix using GANs

[Jimut123]

Types of unsupervised segmentation that can be done in the whole slide:

• Autoencoder based clustering of the cropped grid, and somehow minimize the distance between those data points which are closer and maximize the distance between those data-points that are farther away.

Can also use these:

• PCA based
• Eigen value based
• kNN clustering and grid wise class labels, but this will need a labelled ROI dataset first.

[heraldofsolace]

• kNN clustering and grid wise class labels, but this will need a labelled ROI dataset first.

Can we use the mask I generated?

[Jimut123]

• kNN clustering and grid wise class labels, but this will need a labelled ROI dataset first.

Can we use the mask I generated?

That would be great if we could get the texture of the cell, and it would do something like averaging of textures and find the grids which has the similar kind of texture as the cell. So, we need tight bounding boxes or it would be great if we could only get the mask with a black background, which is easy to generate.

[heraldofsolace]

if we could only get the mask with a black background,

It is a mask with black background

[Jimut123]

Yeah now and this with the original image

[Jimut123]

I mean you get it right?, just the part of the image which has the mask and the rest is black background

[heraldofsolace]

Yes Bubba

[heraldofsolace]

https://drive.google.com/file/d/1sJR6luureEUpdpssG3mut9XRvqnlIduw/view?usp=sharing

[Jimut123]

How many files are there btw?

[heraldofsolace]

Wait I forgot the other ones

[Jimut123]

You know how powerful this data-set is ? which you have created... We can easily train an adversarial learner to detect and segment cell in the wild...

https://arxiv.org/pdf/1312.6082.pdf

[Jimut123]

I think it has the mask + the image... the color is too pinkish to violet... Am I right? We just need to image + black background... not the mask...

[heraldofsolace]

I don't think so. Which image?

[Jimut123]

Could I see side by side the original image (I mean the very first image) and a corresponding image from the data-set (the above sample type) which you created.

[heraldofsolace]

[Jimut123]

Oh cool thanks, then it's okay....

Wait I forgot the other ones

What ?

[Jimut123]

So now we have to build a GAN which learns the distribution of the color of images... A perfect discriminator which we can reuse for detection / classification / segmentation task.

[heraldofsolace]

What?

I forgot some. The 2nd dataset is complete

[Jimut123]

Cool then send, I'll start it today evening.

[heraldofsolace]

https://drive.google.com/file/d/1uqpNnOLosxr06xId3x2B4eoz70pQq0uH/view?usp=drivesdk

[Jimut123]

Exp - 1 30% success

[Jimut123]

Oie @heraldofsolace, any suggestions how to share over 200 MBs of papers related to this repo?