CPMP: Predicting MammaPrint Recurrence Risk from Breast Cancer Pathological Images Using a Weakly Supervised Transformer
🚀NEWS: [2025-10-22]: The cohort has been deposited in the NGDC, BioProject accession number PRJCA035385.
✨NEWS: Our paper is now avavilable on bioRxiv. [preprint link]
We presented an AI-driven computational pathology framework (CPMP) for MP-informed recurrence risk assessment in patients with early-stage BC using histopathological WSIs.
Motivation
Recurrence of breast cancer (BC) is a primary cause of mortality and is associated with poor prognosis. The genomic MammaPrint (MP) test is designed to stratify risk and evaluate chemotherapy benefits for early-stage HR+/HER2- BC patients. However, it fails to reveal spatial tumor morphology and is limited by high costs.
Hypothesis
We hypothesized that morphological patterns in routine histological slides associated with prognosis can be bridged by the MP molecular biomarker.
Key Ideas & Main Findings
- Data Establishment. We established a HR+/HER2- BC clinical cohort, TJMUCH-MP, with 477 female patients from TJMUCH. The patient cohort was comprised of clinicopathologic characteristics, digital WSIs, and genomic MP diagnostic test results.
- Methodological Innovation. We developed a weakly-supervised learning model CPMP, capable of aggregating morphological and spatial features using the agent-attention transformer, to efficiently predict MP recurrence risk from annotation-free BC pathological slides.
- Knowledge Discovery. We further leverage this artificial intelligence (AI) model for spatial and morphological analysis to explore histological patterns related to MP risk groups. CPMP captures the spatial arrangement of tumors at the whole-slide-level and highlights differences in intercellular interaction patterns associated with MP. We characterized the diversity in tumor morphology, uncovering MP High-specific, Low-specific, and colocalized morphological phenotypes that differs in quantitative cellular compositions.
-
External Validation. We conducted prognostic analysis on the external cohort. CPMP model highlighted superior generalization capabilities on the TCGA-BRCA cohort, exhibiting a significant risk-group stratification on multiple prognostic indicators.
-
Clinical Application. CPMP model has the potential to serve as a flexible supplement integrating into routine clinical diagnostic workflow for early-stage BC patients.
The codes have been implemented on Ubuntu 22.04 LTS. The installation in other operation systems (Windows or MAC) should be cautious due to python packages.
CPMP is implemented by Python 3.8 and PyTorch 2.2. The packages required have been provided in the file environments.yml
✨ To install the Conda environment, you can run the command in the terminal:
sudo apt install openslide-tools python3-openslide libmapnik-dev mapnik-utils python3-mapnik
conda env create -f environments.yml
conda activate cpmpNOTE: The codes require package openslide python, but its installation is different between Linux and Windows. Please follow the offical documentation to install and import it in python to make sure it can work correctly.
If ctranspath feature extractor is further used (which was utilized for ablation study in our work), timm library should be installed as described in CTransPath.
The clinical cohort has been deposited in the National Genomics Data Center (NGDC) under BioProject accession number PRJCA035385, which is publicly accessible at: https://ngdc.cncb.ac.cn/bioproject/. Researchers can request access to the data for non-profit, purely academic research purposes following the standard approval process.
The TCGA-BRCA data (slides and clinicopathological information) are available at portal.gdc.cancer.gov. The follow-up data are available at TCGA Clinical Data Resource.
You should put whole slide images into this path --> configured parameter: datapath
The CPMP model is based on the weakly supervised multi-instance learning (MIL) framework. It includes:
datapath="/home/cyyan/Projects/TJMUCHMMP-cohort/"
patchpath="results/TJMUCH_MMP_1WsiPatching_20x"
tile_size=512
overlap_size=512
downsample_factor=2 # {tile_size: ds_factor} mapping should be {256: 1, 512: 2, 1024: 4, 2048: 8}
echo "WsiPatching..."
python WsiTiling.py \
-s $datapath \
-d $patchpath \
-ps $tile_size \
-ss $overlap_size \
--patch \
--bgtissue \
--stitchdatapath="/home/cyyan/Projects/TJMUCHMMP-cohort/"
patchpath="results/TJMUCH_MMP_1WsiPatching_20x"
featspath="results/TJMUCH_MMP_2Feats_20x"
tocsvpath=$patchpath"/process_list_autogen.csv"
batchsize=500
echo "FeatsExtraction..."
pretrain_model="uni"
CUDA_VISIBLE_DEVICES=0 python FeatsExtracting.py \
--feat_to_dir $featspath"_"$pretrain_model \
--csv_path $tocsvpath \
--h5_dir $patchpath \
--pretrain_model $pretrain_model \
--slide_dir $datapath \
--slide_ext ".svs" \
--batch_size $batchsize \
--resize_size 224 \
--custom_downsample $downsample_factor \
--float16NOTE 1: Pre-trained UNI foundation model can be found at github.com/mahmoodlab/UNI
NOTE 2: You can also apply more foundation models ('gigapath', 'H-optimus-0', 'phikon', 'virchow', 'ctranspath', ...) as we did in the ablation study. You need to mkdir pretrained_model_weights/ and put these pretrained models into pretrained_model_weights/.
csvinfopath="results/MPcohort_info.csv"
splitspath="results/TJMUCH_MMP_3CaseSplits"
featspath="results/TJMUCH_MMP_2Feats_20x"
labelname="MMPrisk"
echo "Cross Validation splitting ..."
echo "5 times 5 folds cross validation mode split."
python CaseSplitting.py \
--task_name "MMPrisk" \
--csv_info_path $csvinfopath \
--split_to_dir $splitspath \
--times 5 \
--kfold 5 \
--val_frac 0 \
--test_frac 0.2 \
--label_column_name $labelname \
--label_list "Low" "High"\
--slide_featspath $featspath\
--seed 2020NOTE: val_frac is set to 0.0 and test_frac is 0.2. This means that we splitted 20% data for testing and the remained is for kfold cross-validation.
NOTE: you can also execute this bash script for Whole slide images processing, Feature representation, and Patient splitting in one stop.
CUDA_VISIBLE_DEVICES=0 python ModelTraining.py --config cfgs/TJMUCH_MMP.yamlor
sh run_train_MMPrisk.sh # also write train-logs to the local.NOTE 1: The config hyperparameters file cfgs/TJMUCH_MMP.yaml can be refered to HERE
NOTE 2: You can also observe loss curves and indicator changes in Browser (eg. localhost:6006) during model training by tensorboard.
cd ${results_dir}${datasource}${task}${exp_code}
tensorboard --logdir . --port=<your port, eg. 6006>The training scripts above will generate the evaluated quantitative results to the local in the form of csv-format tables.
We also implemented a script for individual quantitative evaluation (as what we did in the efficiency evaluation). You can execute it by modifying the specific config parameters.
python Evaluation.py -m <absolute exp_code path> -n <No. of tiles sampled, eg. 5,000/10,000>A new WSI (such as slides in TCGA-BRCA) can be directly inferred. Also, the attention heatmap results at the whole-slide level will be generated by the CPMP model.
You should set the data_dir to the path of your slide data, refer to more parameter configurations in wsi_heatmap_params.yaml for guidance.
python SlideInference.py --configs cfgs/wsi_heatmap_params.yaml or you can set parameters directly,
python SlideInference.py --configs cfgs/wsi_heatmap_params.yaml --opts \
data_dir <your_slide_rootpath> \
slide_ext <format of slide, eg. .svs> \
ckpt_path <pre-trained model names, eg. xx/CPMP_v1.pt>where the parameter ckpt_path is the path of the trained models, and data_dir represents the path of whole slide images. You can find and download the trained models below.
✨NEWS: We give an demo example for slide inference and visualization. See HERE.
Configured hyper-parameters for model training are listed below:
CKPTS:
datasource: TJMUCH_MMP
task: MMPriskRegression
exp_code: debug
wsi_root_dir: /home/cyyan/Projects/TJMUCHMMP-cohort/ # wsi root path
data_root_dir: results/TJMUCH_MMP_2Feats_20x_uni # features representation data directory for tiles
csv_info_path: results/MPcohort_info.csv # csv file with slide_id and label information
split_dir: results/TJMUCH_MMP_3CaseSplits_20x/MMPrisk_KFoldsCV # casesplitting root path
use_h5: False # use h5 file or not for coords
results_dir: results/developed_models # results directory for model, logs, and test evaluation results
cfg_to_name: params_setting.txt # hyper-params setting file for saving
indep_data_root_dir: # features representation data directory for indepedent test data
indep_csv_info_path: # indepedent test csv file wih slide info and mp label for evaluation
zeroshot_idx_dir: # results/TJMUCH_MMP_2Feats_20x_plip/pt_files_zs_tissueidx # noisy filtering exp. use zeroshot index for training
TRAIN:
model_type: CPMP # (default: CPMP), and comparative models, ['TransMIL', 'CLAM', 'ABMIL', 'Transformer']
encoding_dim: 1024 # patch encoding dim, [1024, 768, 2048]
model_size: uni1024 # size of model, ['resnet_small', 'ViT_small', 'ccl2048', 'gigapath1536']
agent_token_num: # the number of agent tokens in for CPMP, (default: 1), 'None' indicates 'Transformer' model
num_perslide: 5000 # None will cover all samples, for data sampling, __getitem__
label_col: MMP_index # label column name
labels_list: ["Low", "High"]
n_classes: 1 # number of classes (default: 1) for regression or classification
loss_func: MSE # slide-level classification loss function, (default: 'MSE') or 'CE' for classification
loss_func_aux: MSE # slide-level auxiliary loss function
log_data: True # log data using tensorboard
weighted_sample: False # enable weighted sampling
HyperParams:
max_epochs: 1000 # maximum number of epochs to train (default: 1000)
batch_size: 1 # batch size commonly set to 1 for MIL, we utilized Gradient Accumulation below.
gc: 32 # Gradient Accumulation Step. HERE, Gradient Accumulation is equal to common batch size.
lr: 0.0001 # learning rate (default: 0.0001)
optim: adam # optimizer, adam sgd adamW radam Lamb
scheduler: LinearLR # optimizer scheduler [CosineAnnealingLR CyclicLR LinearLR OneCycleLR StepLR]
drop_out: 0.5 # a float num, enable dropout (p=0.5)
early_stopping: True # enable early stopping or not
early_stopping_metric: score # early stopping metric 'score' or 'loss'
reg: 0.0001 # weight decay (default: 1e-4)
lambda_reg: 0.00001 # L1-Regularization Strength (Default 1e-5)
CROSSVALIDATAION:
times: 5 # number of times (default: 5)
t_start: -1 # start time (default: -1, last time)
t_end: -1 # end time (default: -1, first time)
k: 5 # number of folds (default: 5)
k_start: -1 # start fold (default: -1, last fold)
k_end: -1 # end fold (default: -1, first fold)
COMMON:
gpu: '0'
seed: 2020 # random seed for reproducible experiment (default: 2020)
workers: 8 # data loader workersThe pre-trained weights of the CPMP model on our in-house cohort can be downloaded in HERE. The weights can be used to reproduce the results in our paper. Also, It can be directly applied to new H&E WSIs for inference.
If you have any problems, just raise an issue in this repo.
Our code is developed on the top of CLAM, PhiHER2, and Agent-Attention.
If you find our work useful in your research, please consider citing our work and giving this repo a star. Thank you~
@article {Yan2025.04.26.648504,
author = {Yan, Chaoyang and Li, Linwei and Qian, Xiaolong and Ou, Yang and Huang, Zhidong and Ruan, Zhihan and Xiang, Wenting and Liu, Hong and Liu, Jian},
title = {AI-driven computational pathology for recurrence risk assessment in early-stage breast cancer},
elocation-id = {2025.04.26.648504},
year = {2025},
doi = {10.1101/2025.04.26.648504},
publisher = {Cold Spring Harbor Laboratory},
URL = {https://www.biorxiv.org/content/early/2025/04/29/2025.04.26.648504},
eprint = {https://www.biorxiv.org/content/early/2025/04/29/2025.04.26.648504.full.pdf},
journal = {bioRxiv}
}