Information Maximization-Based Clustering of Histopathology Images Using Deep Learning

This is the repository that has the code and figures for the paper titled Information Maximization-Based Clustering of Histopathology Images Using Deep Learning.

128, which is a folder contains models (in folder models14) for 14-cluster set using 128x128 pixels patches, script for training this particular model (part1_kpc_cae_128_14.ipynb), clustering output using this model (part2_clustering.ipynb) and umap plotting (part3_umap.ipynb) using upper-dimensional latent features obtained from this model. Similarly, Folder 64 includes models (in folder models14) for 14-cluster set using 64x64 pixels patches, script for training this particular model (part1_kpc_cae_64_14.ipynb), clustering output using this model (part2_clustering.ipynb) and umap plotting (part3_umap.ipynb) using upper-dimensional latent features attained from this model.

We have only shown these two models because they have been designated as the optimal cluster sets by internal validation metrics. By changing in one place, the scripts can be run for cluster sets using 8, 9, 10, 11, 12, 13, 15, 16, 17 and 18 as well. We mentioned where to change in the training scipts (part1_kpc_cae_128_14.ipynb and part1_kpc_cae_64_14.ipynb). Check that out.

• 'Figures' (another folder) has every figure used in the manuscript (including Supporting information segment). The Supporting information figures shown in the paper can be found in the Additional Information subfolder here. Besides, some high resolution WSIs can also be found in the Figures folder.

• 'MODULE' folder has been used to calculate internal validation indices like Xie-Beni index, Calinski-Harabasz index, C index, Dunn index, Hartigan index and Mclain-Rao index for both patch sizes (128x128 and 64x64 pixels). Go through these two scripts: Internal_validation_128x128.ipynb and Internal_validation_64x64.ipynb to see index calculations.

• internal_validation_indices_plot.ipynb has the code for creating Figure 5 as shown in the manuscript.

sequential_patch_creation_from_1_WSI.ipynb stores the code for creating 18252 patches sequentially (without overlap) of size 128x128 pixels from a randomly selected WSI with 5 stainings. We could not upload the dataset (seq_patches_18252_from_1wsi.npy) in GitHub because it has a size in Gigabytes but it can be found by accessing this link: seq_patches_18252_from_1wsi.npy. In this script, we have visualized the WSI that was used for sequential patch creation with 5 stainings. Additionally, we have shown a patch from this dataset with 5 stainings. A similar procedure can be followed to create 64x64 patches with minor changes in the script but we did not do that in this research.

In sequential_patch_creation_from_1wsi.ipynb, you will notice HE_2, HE_3 and HE_4 besides HE_1. These HE series were taken for alignment and registration of images. The original tissue was sliced into a series of thin slices. Then, they were stained with HE and other staining methods alternately, e.g., HE-MT-HE-CD31-HE-Ki67-HE-CK...

So, all stained images should be sandwiched by two contacts of HE images.

• IM_output_64x64_100_1.ipynb and IM_output_of_unused_samples_during_training.ipynb possess the code to generate Figure 4 and Figure 5, respectively of the Supporting information portion in the manuscript.

• original_transformed_128x128.ipynb has the code to visualize the original and transformed versions of a patch (128x128 pixels). Look over Figure 2 of the Supporting information division in the paper.

• the_learning_curve.ipynb contains the code to create Figure 3 of the Supporting information segment in the paper.

patch_visualization_128x128_64x64.ipynb has the code to create Figure 1 of the manuscript. In addition, we have shown first 20 samples of both datasets (slice128_Block2_20K.npy ---> contains 128x128 pixels patches and 'slice64_Block2_20K.npy' ---> stores 64x64 pixels patches) with 5 stainings. The sizes of slice128_Block2_20K.npy and slice64_Block2_20K.npy are not suitable for being uploaded in GitHub (Gigabyte files).

But, these two datasets can be found in this link: Location of 128x128 and 64x64 patches

• histogram_colormap_WSI.ipynb keeps the code to create Figure 10 of the manuscript.

filename_tif8_block2.txt contains the addresses of 191 WSIs.

• By using random_patch_creation.ipynb script, randomly patches can be created from 191 WSIs. Here, we have presented the code for creating 128x128 patches but with minor changes, 64x64 patches can also be created. We indicated where to change in the script to create 64x64 patches.

You will observe some of the locations in different scipts of this repository that points to the placement of different types of files (either dataset or something else) as shown below:

• /project/dsc-is/mahfujul-r/M/slice128_Block2.20K.ipynb
• models14/model_en_202211280036_4000.ckpt
• models14/model_cl_202211280036_4000.ckpt
• models14/model_de_202211280036_4000.ckpt
• project/dsc-is/mahfujul-r/M/128/14/HHH14 & C14 & D14.csv
• /project/dsc-is/mahfujul-r/M/slice64_Block2.20K.ipynb
• models14/model_encoder_3000
• models14/model_classifier_3000
• models14/model_decoder_3000
• project/dsc-is/mahfujul-r/M/64/14/HHH14 & C14 & D14.csv
• ./64/models14/model_encoder_3000
• ./64/models14/model_classifier_3000
• ./64/models14/model_decoder_3000
• ./128/models14/model_en_202211280036_4000.ckpt
• ./128/models14/model_cl_202211280036_4000.ckpt
• ./128/models14/model_de_202211280036_4000.ckpt
• project/dsc-is/mahfujul-r/M/128/08/HHH08 & C08 & D08.csv
• project/dsc-is/mahfujul-r/M/128/09/HHH09 & C09 & D09.csv
• project/dsc-is/mahfujul-r/M/128/10/HHH10 & C10 & D10.csv
• project/dsc-is/mahfujul-r/M/128/11/HHH11 & C11 & D11.csv
• project/dsc-is/mahfujul-r/M/128/12/HHH12 & C12 & D12.csv
• project/dsc-is/mahfujul-r/M/128/13/HHH13 & C13 & D13.csv
• project/dsc-is/mahfujul-r/M/128/15/HHH15 & C15 & D15.csv
• project/dsc-is/mahfujul-r/M/128/16/HHH16 & C16 & D16.csv
• project/dsc-is/mahfujul-r/M/128/17/HHH17 & C17 & D17.csv
• project/dsc-is/mahfujul-r/M/128/18/HHH18 & C18 & D18.csv
• project/dsc-is/mahfujul-r/M/64/08/HHH08 & C08 & D08.csv
• project/dsc-is/mahfujul-r/M/64/09/HHH09 & C09 & D09.csv
• project/dsc-is/mahfujul-r/M/64/10/HHH10 & C10 & D10.csv
• project/dsc-is/mahfujul-r/M/64/11/HHH11 & C11 & D11.csv
• project/dsc-is/mahfujul-r/M/64/12/HHH12 & C12 & D12.csv
• project/dsc-is/mahfujul-r/M/64/13/HHH13 & C13 & D13.csv
• project/dsc-is/mahfujul-r/M/64/15/HHH15 & C15 & D15.csv
• project/dsc-is/mahfujul-r/M/64/16/HHH16 & C16 & D16.csv
• project/dsc-is/mahfujul-r/M/64/17/HHH17 & C17 & D17.csv
• project/dsc-is/mahfujul-r/M/64/18/HHH18 & C18 & D18.csv
• ./64/models14/hist_modelS_3000

You have to change these locations according to your working environment.

This repository uses MIT License. Read the terms and conditions from LICENSE text file.

Read the paper from here: KPC_IM