Reconsidering Explicit Longitudinal Mammography Alignment for Enhanced Breast Cancer Risk Prediction

This is the code for our MICCAI 2025 paper. For more details, please refer to our paper. Preprint available at: https://arxiv.org/abs/2506.19363

Table of Contents

1. Introduction
2. Method
3. Datasets
4. Results / Key findings of the paper
5. Reproduction of the results
6. Citation

Introduction

Risk prediction estimates the probability of developing breast cancer within a specified time frame, focusing on a 5-year period following a specific mammogram. However, effective use of longitudinal mammograms requires precise spatial alignment or registration, as misalignment can obscure abnormalities and distort risk assessments. This is challenging due to the non-rigid nature of breast tissue and variations in breast compression during examinations. To address the lack of insights into the choice of alignment, this study aims to evaluate explicit alignment strategies for longitudinal mammography and their impact on risk prediction. We examine whether alignment should occur at the image or feature level. To facilitate efficient image level alignment, we introduce MammoRegNet, the first deep learning-based image registration model for aligning longitudinal mammography images.

Method

We compare different explicit alignment strategies to assess their impact on risk prediction performance

1) NoAlign: Direct feature extraction without alignment

2) FeatAlign: Feature-level alignment (jointly training of alignment and risk prediction)

3) FeatAlignReg: Similar to FeatAlign, we only extend the alignment Loss with a regularization loss

4) ImgAlign: Image-level alignment with MammoRegNet (pretrained MammoRegNet)

5) ImgFeatAlign: Applying MammoRegNet’s deformation field in feature space (pretrained MammoRegNet)

(a) NoAlign	(b) FeatAlign
(c) ImgAlign	(d) ImgFeatAlign

Datasets

We used two large, publicly available mammography datasets :

• Emory Breast Imaging Dataset (EMBED): https://aws.amazon.com/marketplace/pp/prodview-unw4li5rkivs2#overview}
• Cohort of Screen-Aged Women Case Control (CSAW-CC): https://snd.se/en/catalogue/dataset/2021-204-1

Results / Key findings of the paper

Effect of Feature- vs. Image-Level Alignment on Risk Prediction:

• We assess the impact of feature-level and image-level alignment on risk prediction performance

  → Image-level alignment is superior to representation-level alignment leading to better risk prediction performance

1–5 Year Breast Cancer Risk Prediction of different alignment methods: C-index and AUC with 95% CI in parentheses.

In-Depth Analysis of Weighting and Alignment:

• We further explore how different weightings of the alignment loss in the jointly training of feature alignment and risk prediction (in FeatAling) affect risk prediction performance

  → Increasing the alignment weight improves deformation field quality by reducing negative Jacobian determinants but decreases risk prediction performance. This shows that there is a trade-off between alignment quality and predictive performance

• We also evaluate feature-level and image-level alignment across breast density categories

  → Image-level alignment (ImgAlign) performs best across the different density categories especially for dense breast tissue

Analysis of Weighting and Alignment

Reproduction of the results

For reproducing the results follow the instructions below:

Important: for each script in the scripts folder, make sure you update the paths to load the correct datasets and export the results in your favorite directory.

1) Requirements

Requirements are in the requirements.txt file

2) Pre-processing of the datasets

The preprocessing step ensures that the datasets are properly prepared before training.

The preprocessing folder contains the necessary scripts to preprocess images and split the datasets into training, validation and test.

To preprocess the EMBED dataset, use: preprocessing/preprocess_img_embed.py

To preprocess the CSAW-CC dataset, use: preprocessing/preprocess_img_csaw_cc.py

To split both datasets into training, validation, and test sets, use: preprocessing/split_data.py

For the risk prediction, create a CSV file describing your dataset by running the notebooks in the notebooks folder

3) Training

3.1) MammoRegNet

For training MammoRegNet run scripts/train_MammoRegNet.sh

3.2) Risk prediction models

For training the risk prediction models run scripts/train_risk_NoAlign.sh, scripts/train_risk_FeatAlign.sh, scripts/train_risk_FeatAlignReg.sh, scripts/train_risk_ImgAlign.sh, and scripts/train_risk_ImgFeatAlign.sh

4) Inference

4.1) MammoRegNet

Run scripts/test_MammoRegNet.sh

4.2) Risk prediction models

Run scripts/test_risk_NoAlign.sh, scripts/test_risk_FeatAlign.sh, scripts/test_risk_FeatAlignReg.sh, scripts/test_risk_ImgAlign.sh, and scripts/test_risk_ImgFeatAlign.sh

Citation

If you use this code or models in your scientific work, please kindly cite our paper:

@inproceedings{,
  title={Reconsidering Explicit Longitudinal Mammography Alignment for Enhanced Breast Cancer Risk Prediction},
  author={Thrun, Solveig and Hansen, Stine and Sun, Zijun and Blum, Nele and Salahuddin A., Suaiba  and Wickstrøm, Kristoffer and Wetzer, Elisabeth and Jenssen, Robert and Stille, Maik and Kampffmeyer, Michael},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={ },
  year={2025},
  organization={Springer}
}