MTP-BCR: Predicting long term breast cancer risk using longitudinal mammographic screening history

Risk assessment of breast cancer (BC) seeks to enhance individualized screening and prevention strategies. Recent deep learning (DL) risk models based on mammography have shown superiority in short-term risk prediction compared to traditional risk factor-based models. However, those models primarily rely on single-time exams and thus ignore the temporal changes in breast tissues that can be extracted from sequence exams. Here, we present the Multi-Time Point Breast Cancer Risk Model (MTP-BCR), a novel temporospatial DL risk model that integrates traditional BC risk factors and longitudinal mammography data to identify subtle changes in breast tissue indicative of future malignancy. Utilizing a large inhouse dataset comprising risk factors and 171,168 mammograms involving 9,133 women, we evaluate the performance of the MTP-BCR model in long-term risk prediction. Our model demonstrates a significant improvement in 10-year risk prediction with an area under the receiver operating characteristics (AUC) of 0.80, outperforming the traditional BCSC 10-year risk model, our pure image model (without risk factors), and is also superior to other SOTA methods at 5-year AUC in various screening cohorts. Furthermore, MTP-BCR provides unilateral breast-level predictions with AUCs up to 0.81 and 0.77 for 5-year and 10-year risk assessments, respectively. The validation on the public CSAW-CC dataset demonstrates the consistent advantage of our multi-time point-based model compared to the single-time point-based method. For the prediction of breast cancer recurrence, our MTP-BCR obtains a 5-year AUC of 0.71 which also surpasses other methods. The heatmaps derived from our model may help clinicians better understand the progression from normal tissue to cancerous growth, enhancing interpretability in breast cancer risk assessment.

Overview of study

Model architecture

Usage

Requirements:

• pytorch
• numpy
• pandas
• scikit-learn
• ...

Environment

conda env create -f py37_environment.yml  # don't forget to change the path
conda activate py37

Example training:

python src/train.py \
  --method side_specific_4views_mtp_tumor \
  --num-classes 16 \   #   leveraged the 15 years follow-up to train the risk model
  --test-num-classes 11 \  #  test the model's performance of 10-year risk prediction
  --batch-size 8 \
  --num_time_points 6 \  # input 6 time point exams, for one current exams and five history exams
  --use_risk_factors \
  --multi_tasks \
  --pooling last_timestep \
  --years_at_least_followup 0 \
  --results-dir ./log/Mammo_risk/

The configs above are meant to specify exact implementation details and our experimental procedure

Predicting on External Dataset

python custom_predict.py \
    --path_risk_model $dir_pretrained_weigths \
    --test_results_dir $dir_results

Weights

We are going to share the weights by to this link: https://drive.google.com/drive/folders/1VFA1lvihPtTWlQVXbBr_Jot0ugZUZLXL?usp=sharing

Disclaimer

This code and accompanying pretrained models are provided with no guarantees regarding their reliability, accuracy or suitability for any particular application and should be used for research purposes only. The models and code are not to be used for public health decisions or responses, or for any clinical application or as a substitute for medical advice or guidance.

License

The code is MIT licensed, as found in the LICENSE file

Contact details

If you have any questions please contact us.

Email: [email protected] (Ritse Mann); [email protected] (Tao Tan); [email protected] (Xin Wang)

Links: Netherlands Cancer Institute, Radboud University Medical Center and Maastricht University