Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Ordinal Learning: Longitudinal Attention Alignment Model for Predicting Time to Future Breast Cancer Events from Mammograms

Xin Wang, Tao Tan, Yuan Gao, Eric Marcus, Luyi Han, Antonio Portaluri, Tianyu Zhang, Chunyao Lu, Xinglong Liang, Regina Beets-Tan, Jonas Teuwen, Ritse Mann

TL;DR

The paper tackles the challenge of predicting both breast cancer risk and the time to future cancer events from longitudinal mammograms. It introduces OA-BreaCR, an ordinal-learning framework that jointly models time-to-event and risk while explicitly aligning attention between prior and current mammograms to capture tissue changes over time. The approach combines Mean-Variance loss and Probabilistic Ordinal Embedding to enforce ordinal structure and distributional plausibility, plus an attention-alignment module to estimate deformable changes between time points. Experiments on the EMBED and Inhouse datasets show improved time-to-event accuracy and risk stratification, with interpretable attention heatmaps highlighting relevant temporal tissue changes. These results suggest that ordinal, interpretable modeling of longitudinal mammography data can enhance both precision and clinical trust in breast cancer screening and prevention.

Abstract

Precision breast cancer (BC) risk assessment is crucial for developing individualized screening and prevention. Despite the promising potential of recent mammogram (MG) based deep learning models in predicting BC risk, they mostly overlook the 'time-to-future-event' ordering among patients and exhibit limited explorations into how they track history changes in breast tissue, thereby limiting their clinical application. In this work, we propose a novel method, named OA-BreaCR, to precisely model the ordinal relationship of the time to and between BC events while incorporating longitudinal breast tissue changes in a more explainable manner. We validate our method on public EMBED and inhouse datasets, comparing with existing BC risk prediction and time prediction methods. Our ordinal learning method OA-BreaCR outperforms existing methods in both BC risk and time-to-future-event prediction tasks. Additionally, ordinal heatmap visualizations show the model's attention over time. Our findings underscore the importance of interpretable and precise risk assessment for enhancing BC screening and prevention efforts. The code will be accessible to the public.

Ordinal Learning: Longitudinal Attention Alignment Model for Predicting Time to Future Breast Cancer Events from Mammograms

TL;DR

The paper tackles the challenge of predicting both breast cancer risk and the time to future cancer events from longitudinal mammograms. It introduces OA-BreaCR, an ordinal-learning framework that jointly models time-to-event and risk while explicitly aligning attention between prior and current mammograms to capture tissue changes over time. The approach combines Mean-Variance loss and Probabilistic Ordinal Embedding to enforce ordinal structure and distributional plausibility, plus an attention-alignment module to estimate deformable changes between time points. Experiments on the EMBED and Inhouse datasets show improved time-to-event accuracy and risk stratification, with interpretable attention heatmaps highlighting relevant temporal tissue changes. These results suggest that ordinal, interpretable modeling of longitudinal mammography data can enhance both precision and clinical trust in breast cancer screening and prevention.

Abstract

Precision breast cancer (BC) risk assessment is crucial for developing individualized screening and prevention. Despite the promising potential of recent mammogram (MG) based deep learning models in predicting BC risk, they mostly overlook the 'time-to-future-event' ordering among patients and exhibit limited explorations into how they track history changes in breast tissue, thereby limiting their clinical application. In this work, we propose a novel method, named OA-BreaCR, to precisely model the ordinal relationship of the time to and between BC events while incorporating longitudinal breast tissue changes in a more explainable manner. We validate our method on public EMBED and inhouse datasets, comparing with existing BC risk prediction and time prediction methods. Our ordinal learning method OA-BreaCR outperforms existing methods in both BC risk and time-to-future-event prediction tasks. Additionally, ordinal heatmap visualizations show the model's attention over time. Our findings underscore the importance of interpretable and precise risk assessment for enhancing BC screening and prevention efforts. The code will be accessible to the public.
Paper Structure (11 sections, 3 figures, 2 tables)

This paper contains 11 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Definition of the risk and time-to-BC prediction tasks
  4. Ordinal learning for estimating the time to future BC events
  5. Risk-related attention alignment of prior-current mammograms
  6. Experimental
  7. Experimental settings
  8. Experimental results
  9. Conclusion
  10. Acknowledgments.
  11. Disclosure of Interests.

Figures (3)

  • Figure 1: The schematic overview of the proposed OA-BreaCR. (A) Our method utilizes ordinal learning and longitudinal attention alignment for the risk and time-to-BC prediction tasks. (B) Ordinal learning for estimating the time to future BC events by leveraging Mean-variance (MV) Loss pan2018mean and probabilistic ordinal embedding (POE) li2021learning. (C) The attention alignment model aims to learn temporal breast tissue changes from two-time point mammograms in an interpretable manner.
  • Figure 2: Heatmaps of attention alignment on prior-current MG. The red circles indicate the the existing tumors or high-risk areas that developed the cancer within five years.
  • Figure 3: Violin plots show the distributions of the expected time to cancer in stratified three high-risk populations. All dots indicate true high-risk patients whom the risk model successfully identified. Colored dots represent patients identified by the risk model with an estimated time to cancer within 5 years. While grey dots represent cases where the estimated time-to-cancer exceeds five years, indicating failures to accurately estimate time-to-cancer. n is the number of color dots, with the total count of color and grey dots provided in parentheses.