Skip to content

balaganeshmohan/Colorectal-cancer-classification-with-histopathological-images

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

22 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Detection-of-tumor-mutation-in-histopathological-images-colorectal-cancer

This project involves a novel way to train CNN's for the purposes of detecting mutation in colorectal cancer cells. The pipiline involves training 3 different architectures each with 3 models of different resolutions, specifically 128x128, 256x256 and 512x512. Each model feeds their weight to a larger model similar to transfer learning. So in our case, 128x128--> 256x256 --> 512x512. The reason to do this is to overcome invariance in CNN, as the process involves cutting the original tumour tissue of a patient into several tiles because of memory restrictions to train on a GPU.

The images in this repo seems to break when night mode is ON.

Data

The training data was downloaded from here. An example batch of training data at native resolution is shown below. MSIH is the tumour positive class and nonMSIH is tumour negative class.

Screenshot

Training pipeline

There are 3 different models of different architecture and each architecture is trained on different resolutions of cancer images and their are progressively fed from smaller model to the largest being the final predictive model. So in total we have 3 predictive models of which they are ensembled to produce the final results.

Screenshot

Results of the final ensembled models

Auc Score

Screenshot

PR Curve

Screenshot

Final predictions

The final predictions was done on 512x512 images. With a total 144 patients and each patient averaging 225 slides(images), a prediction was done on each slide and for the final predicted class the median of all the probabilities of the individual tiles are calculated.

Screenshot

All code can be seen in this notebook .

Plotting the cooridnates of mutation

It would be helpful to see if there are any spatial patterns in which the mutation occurs in this type of cancer by plotting the predicted class of the feature set. I was not able to find any particular pattern from analyzing all the different patients in the test set. Screenshot

Screenshot

Acknowledgements

Professor Rachel Cavill - Maastricht University

Professor Jakob Nicholas - Aachen University

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published