PANTHER

Morphological Prototyping for Unsupervised Slide Representation Learning in Computational Pathology, CVPR 2024.
Andrew H. Song*, Richard J. Chen*, Tong Ding, Drew F.K. Williamson, Guillaume Jaume, Faisal Mahmood

Arxiv | Cite

Abstract: Representation learning of pathology whole-slide images (WSIs) has been has primarily relied on weak supervision with Multiple Instance Learning (MIL). However, the slide representations resulting from this approach are highly tailored to specific clinical tasks, which limits their expressivity and generalization, particularly in scenarios with limited data. Instead, we hypothesize that morphological redundancy in tissue can be leveraged to build a task-agnostic slide representation in an unsupervised fashion. To this end, we introduce PANTHER, a prototype-based approach rooted in the Gaussian mixture model that summarizes the set of WSI patches into a much smaller set of morphological prototypes. Specifically, each patch is assumed to have been generated from a mixture distribution, where each mixture component represents a morphological exemplar. Utilizing the estimated mixture parameters, we then construct a compact slide representation that can be readily used for a wide range of downstream tasks. By performing an extensive evaluation of PANTHER on subtyping and survival tasks using 13 datasets, we show that 1) PANTHER outperforms or is on par with supervised MIL baselines and 2) the analysis of morphological prototypes brings new qualitative and quantitative insights into model interpretability.

Updates

• 07/02/2024: Multimodal extension of PANTHER (MMP) codebase is now live.
• 05/06/2024: The first version of PANTHER codebase is now live.

Installation

Once you clone the repo, please run the following command to create PANTHER conda environment.

conda env create -f env.yaml

PANTHER Walkthrough

There are two workflows for PANTHER, depending on the use case.

• Workflow 1
  • Step 0 ⇒ Step 1 ⇒ Step 2A ⇒ (Step 3)
  • Useful for constructing unsupervised slide representations, without any specific downstream tasks at hand.
• Workflow 2
  • Step 0 ⇒ Step 1 ⇒ Step 2B ⇒ (Step 3)
  • Useful for when there is a specific downstream task at hand. The slide representations will be constructed.

Step 0. Dataset organization

Data csv: The data csv files (with appropriate splits, e.g., train, test) are placed within src/splits with appropriate folder structure. For example, for classification task on ebrains, we would have

splits/
	├── ebrains
    		├── train.csv
    		├── val.csv
    		└── test.csv

Alternatively, for 5-fold cross-validation survival task on TCGA BRCA, we would have

splits/
	├── TCGA_BRCA_survival_k=0
    		├── train.csv
    		├── val.csv
    		└── test.csv
	├── ...

        ├── TCGA_BRCA_survival_k=4
    		├── train.csv
    		├── val.csv
    		└── test.csv

Patch features: For the following steps, we assume that features for each patch have already been extracted and that each WSI is represented as a set of patch features. For examples of patch feature extraction, please refer to CLAM.

The code assumes that the features are either in .h5 or .pt formats - the feature directory path FEAT_DIR has to end with the directory feats_h5/ if the features are in .h5 format, and feats_pt/ for .pt format.

While there is no de facto standard, one good practice of organizing features are as follows (used as examples in clustering and panther)

/path_to_data_folder/tcga_brca/extracted_mag20x_patch256_fp/extracted-vit_large_patch16_224.dinov2.uni_mass100k/feats_h5

which specifies magnification, patch size, and feature extractor used to create the patch features.

Step 1. Prototype construction

For prototype construction, we use K-means clustering across all training WSIs. We recommend using GPU-based FAISS when using large number of patch features for clustering. For example, we can use the following command to find 16 prototypes (of 1,024 dimension each) using FAISS from WSIs corresponding to SPLIT_DIR/train.csv.

CUDA_VISIBLE_DEVICES=0 python -m training.main_prototype \
--mode faiss \
--data_source FEAT_DIR_1,FEAT_DIR_2 \
--split_dir SPLIT_DIR \
--split_names train \
--in_dim 1024 \
--n_proto_patches 1000000 \
--n_proto 16 \
--n_init 5 \
--seed 1 \
--num_workers 10 \

The list of parameters is as follows:

• mode: 'faiss' uses GPU-enabled K-means clustering to find the prototypes. 'kmeans' uses sklearn K-means clustering on CPU ('faiss' or 'kmeans').
• data_source: comma-separated list of feature directories ending with either feats_h5 or feats_p5. Example of a feature dictory is provided in Step 0.
• split_names: Which data split to perform clustering/prototyping on. By default train is the best (Since train split has the most data.)
• in_dim: Dimension of the patch features, dependent on the feature encoder.
• n_proto: Number of prototypes.
• n_proto_patches: Number of patch features to use per prototype. In total, n_proto * n_proto_patches features are used for finding prototypes.
• n_init: Number of K-means initializations to try.

The prototypes will be saved in the SPLIT_DIR/prototypes folder.

A concrete script example of using TCGA-BRCA patch features can be found below.

cd src
./scripts/prototype/brca.sh 0

This will initiate the script scripts/prototype/clustering.sh for K-means clustering. Detailed explanations for clustering hyperparameters can be found in clustering.sh.

Step 2A. Unsupervised slide representation construction

Once the prototypes are constructed, we can use PANTHER or OT to construct unsupervised slide representations.

cd src
./scripts/embedding/brca.sh 0 panther

This step will create two files in the SPLIT_DIR/embeddings folder: 1) (*.pkl) original slide-level representation and 2) (*_tokenized.pkl) slide-level representation tokenized into each prototype and statistics (e.g., mixture probability, mean, cov). Note that for OT, mixture probability is uniform and cov=None.

Alternatively, if you want to construct slide representations as part of a classification or survival downstream task, you can skip Step 2A and go straight to Step 2B.

Step 2B. Training downstream model

Once the prototypes are constructed in Step 1 (Step 2A is not required), we can run a downstream task

cd src
./scripts/survival/brca_surv.sh 0 panther

Step 3. Visualization

To visualize GMM mixture proportions in prototypical assignment maps in PANTHER, see the accompanying notebook.

Additional Findings in PANTHER

• We observe that using high-quality pretrained ROI encoders (such as UNI) leads to significant performances across all MIL and set-based learning methods (see the Supplement). Specifically:
• • When using Resnet-50 (ImageNet Transfer) and CTransPath features, unsupervised set representation methods such as OT and PANTHER underperform MIL methods (using the same features). With UNI features, OT and PANTHER can readily outperform MIL, and should be considered strong baselines when evaluating slide-level tasks.
• • DeepAttnMISL with UNI features becomes a strong MIL baseline. This can be attributed to DeepAttnMISL dependent on K-Means for cluster pooling (which depends on high-quality representations).
• • With unsupervised slide representations extracted per WSI (via OT or PANTHER), training survival models on WSIs is now much more stable since you can directly use Cox loss (instead of NLL). Across all of our ablation experiments, PANTHER with UNI features always achieved C-Index > 0.6.

PANTHER Limitations

As unsupervised slide representations in PANTHER are created using non-parametric techniques such as K-Means Clustering and GMMs (which rely on Euclidean distance or dot product to compare embeddings), we note the following limitations:

• Dependent on the degree of dataset shift between the train and test distributions (due to variable H&E stain variability, known as image acquisition shift), prototype assignment for certain WSIs may lead to results in which all patches are assigned to a single prototype. This is exemplified in TCGA which has site-specific biases, and is thus an important consideration when considering using PANTHER (or any non-parametric approach) for histopathologic biomarker discovery.
• When clustering over a WSI dataset composed of millions to billions of patches, clustering with only C=16 clusters will likely underfit the dataset, and also lead to collapse of all patches in a WSI falling under a single prototype. Empirically, we found C=16 to outperform C=32 in supervised settings. However, in settings such as biomarker discovery or unsupervised tissue segmentation, using more prototypes may improve performance.

Acknowledgements

If you find our work useful in your research or if you use parts of this code please cite our paper:

@inproceedings{song2024morphological,
    title={Morphological Prototyping for Unsupervised Slide Representation Learning in Computational Pathology},
    author={Song, Andrew H and Chen, Richard J and Ding, Tong and Williamson, Drew FK and Jaume, Guillaume and Mahmood, Faisal},
    booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
    year={2024},
}

The code for PANTHER was adapted and inspired by the fantastic works of DIEM, OTK, and CLAM. Boilerplate code for setting up supervised MIL benchmarks was developed by Ming Y. Lu.

Issues

• Please open new threads or report issues directly (for urgent blockers) to asong@bwh.harvard.edu.
• Immediate response to minor issues may not be available.