Interpretable artificial intelligence-based determination of glioma IDH mutation status directly from histology slides

Isocitrate dehydrogenase (IDH) mutation status is a diagnostic requirement for glioma with associated prognostic and therapeutic implications. Clinical routine visual assessment of tissue is insufficient to determine IDH status conclusively, mandating molecular workup that is unavailable everywhere. We developed an interpretable Artificial Intelligence (AI)-based approach for determining IDH status directly from H&E-stained glioma slides.

Environment

Pre-requisites:

• Linux (Tested on Ubuntu 22.04)
• NVIDIA GPU (Tested on Nvidia A6000)

After setting up anaconda, first install openslide:

sudo apt-get install openslide-tools

Next, use the environment configuration file to create a conda environment:

conda env create -n idh_classifier -f  idh_classifier.yaml

Activate the environment:

conda activate idh_classifier

Once inside the created environment, to install smooth-topk (first cd to a location that is outside the project folder and is suitable for cloning new git repositories):

git clone https://github.com/oval-group/smooth-topk.git
cd smooth-topk
python setup.py install

WSI Segmentation and Patching

data/slides_20x/
	├── patient_1_slide_a.svs
	├── patient_1_slide_b.svs
	└── ...
data/slides_40x/
	├── patient_2_slide_a.svs
	├── patient_2_slide_b.svs
	└── ...

🛠 Workflow Logic

The pipeline automatically adjusts the extraction scale and file paths based on the input magnification. This ensures that the physical area covered by a patch remains consistent or follows your specific protocol.

Input Argument	Target Data Directory	Output Directory	Patch Size
20x	data/slides_20x/	data/slides_patches_20x/	256
40x	data/slides_40x/	data/slides_patches_40x/	512

Execution Examples

# Process 20x slides with 256px patches
./create_patches.sh 20x

# Process 40x slides with 512px patches
./create_patches.sh 40x

By setting 20x to 256 and 40x to 512, you are effectively keeping the field of view (FOV) of each patch identical in terms of physical microns (assuming the 40x scan has twice the resolution of the 20x scan). This is a standard best practice in pathology machine learning to ensure the model sees the same amount of tissue per tile regardless of the scanner settings.

Output Directory Structure

data/slides_patches_20x/
	├── masks
    		├── patient_1_slide_a.png
    		├── patient_1_slide_b.png
    		└── ...
	├── patches
    		├── patient_1_slide_a.h5
    		├── patient_1_slide_b.h5
    		└── ...
	├── stitches
    		├── patient_1_slide_a.png
    		├── patient_1_slide_b.png
    		└── ...
	└── slides_processed.csv

data/slides_patches_40x/
├── masks
        ├── patient_1_slide_a.png
        ├── patient_1_slide_b.png
        └── ...
├── patches
        ├── patient_1_slide_a.h5
        ├── patient_1_slide_b.h5
        └── ...
├── stitches
        ├── patient_1_slide_a.png
        ├── patient_1_slide_b.png
        └── ...
└── slides_processed.csv

🧹 Patch Cleanup (Step 2)

After initial patching, the pipeline runs a Cleanup Script to filter out low-quality tiles.

Filtering Criteria:

1. White Space: Patches with >85% background are removed.
2. Stain Detection: Uses HED (Hematoxylin-Eosin-DAB) color deconvolution to ensure tissue is actually present.
3. HSV Filtering: Removes blurry or out-of-focus areas based on saturation and value thresholds.

Why this is necessary:

Whole Slide Images often contain artifacts, marker ink, or large empty regions. By cleaning the .h5 files, you reduce the noise in your training set and significantly speed up the feature extraction (encoding) step.

Creating Features

Run the extraction script by specifying magnification, batch size, and the desired model backbone. The script dynamically maps to the correct data and coordinate directories based on the magnification provided.

Usage

./extract_features.sh <MAG> <BATCH_SIZE> <BACKBONE>

Example:

./extract_features.sh 20x 256 uni

Supported Backbones

We support several state-of-the-art self-supervised and supervised models for histopathology.
For more details about each model, please refer to the original repositories to request access and follow their specific licensing terms.

• UNI : https://github.com/mahmoodlab/UNI
• CTransPath : https://github.com/Xiyue-Wang/TransPath
• RetCCL : https://github.com/Xiyue-Wang/RetCCL
• HIPT-4K : https://github.com/mahmoodlab/HIPT
• Lunit ViT : https://github.com/lunit-io/benchmark-ssl-pathology
• SimCLR : https://github.com/ozanciga/self-supervised-histopathology
• ResNet-50 : ImageNet pretrained

Training the models

Usage Instructions

To run the training script, pass the magnification level and backbone name as arguments:

chmod +x train.sh
./train.sh <MAG> <BACKBONE>

Example:

./train.sh 20x uni

Methodology

This codebase is heavily based on CLAM. However, unlike CLAM, instance-level clustering is not used. The model is trained using a pure Attention-based Multiple Instance Learning (MIL) framework.

As we discussed in our paper, we have trained our model on three different groups of patients. To train the train model replace the splits_dir argument in the main.py file.

Evaluation

To run the evaluation script, pass the magnification level and backbone name as arguments:

chmod +x eval.sh
./eval.sh <MAG> <BACKBONE>

Example:

./eval.sh 20x uni

References

1. Lu, Ming Y., et al. "Data-efficient and weakly supervised computational pathology on whole-slide images." Nature biomedical engineering 5.6 (2021): 555-570.
2. Wang, Xiyue, et al. "RetCCL: Clustering-guided contrastive learning for whole-slide image retrieval." Medical image analysis 83 (2023): 102645.
3. Kang, Mingu, et al. "Benchmarking self-supervised learning on diverse pathology datasets." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.
4. Chen, Richard J., et al. "Towards a general-purpose foundation model for computational pathology." Nature medicine 30.3 (2024): 850-862.
5. Wang, Xiyue, et al. "Transformer-based unsupervised contrastive learning for histopathological image classification." Medical image analysis 81 (2022): 102559.
6. Srinidhi, Chetan L., and Anne L. Martel. "Improving self-supervised learning with hardness-aware dynamic curriculum learning: an application to digital pathology." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021.
7. Chen, Richard J., et al. "Scaling vision transformers to gigapixel images via hierarchical self-supervised learning." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022.

Citations

Shubham Innani, W Robert Bell, Hannah Harmsen, MacLean P Nasrallah, Bhakti Baheti, Spyridon Bakas, Interpretable artificial intelligence based determination of glioma IDH mutation status directly from histology slides, Neuro-Oncology Advances, Volume 7, Issue 1, January-December 2025, vdaf140, https://doi.org/10.1093/noajnl/vdaf140

@article{10.1093/noajnl/vdaf140,
    author = {Innani, Shubham and Bell, W Robert and Harmsen, Hannah and Nasrallah, MacLean P and Baheti, Bhakti and Bakas, Spyridon},
    title = {Interpretable artificial intelligence based determination of glioma IDH mutation status directly from histology slides},
    journal = {Neuro-Oncology Advances},
    volume = {7},
    number = {1},
    pages = {vdaf140},
    year = {2025},
    month = {07},
    issn = {2632-2498},
    doi = {10.1093/noajnl/vdaf140},
    url = {https://doi.org/10.1093/noajnl/vdaf140},
    eprint = {https://academic.oup.com/noa/article-pdf/7/1/vdaf140/63738190/vdaf140.pdf},
}