Temporal Mammogram Machine Learning for Breast Cancer Risk Prediction with Uncertainty-Quantified Decision Referrals Open Access

Abstract

Breast cancer is one of the most common diseases and leading causes of mortality in women. Radiologists increasingly use AI systems as a second opinion to enhance screening accuracy from mammography exams. However, most deep learning models rely on a single scan and overlook longitudinal comparisons routinely used in clinical practice. A key challenge in using prior mammograms is that patient positioning and compression can make temporal comparisons unreliable without anatomical alignment. Furthermore, existing models rarely communicate predictive uncertainty, which may encourage over-reliance on outputs even in ambiguous cases. Limited research has explored approaches that include both patients' complete longitudinal mammogram scans and precisely spatially aligning these scans, as well as comparing and evaluating the contributions of each to predictive performance. This thesis conducts a comprehensive comparison of mammography screening models with multiple priors, spatial alignment, and uncertainty quantification. Using the EMBED Open dataset, we trained models that predict the risk of breast cancer developing within one to five years by combining a patient’s current mammogram with all available prior scans, spatially aligning these exams over time, and quantifying model uncertainty using entropy-based measures. The models refer uncertain cases to clinician for additional review. We compare the performance of five deep learning models to answer: (i) does temporal modeling improve future cancer risk predictions; (ii) does temporal alignment improve future cancer risk predictions; and (iii) do uncertainty quantified decision referrals help with clinician decision making? Our results demonstrate that spatially aligning a prior exam yields a statistically significant improvement (p < 0.05) in precision-recall performance (AUPRC) across all prediction horizons compared to using an unaligned prior exam. Temporal models with aligned priors also showed favorable point estimates relative to the current-exam baseline, although these differences were not statistically significant on this dataset. Incorporating multiple priors did not consistently improve performance over a single aligned prior. The results suggest that spatial alignment helps capture more meaningful longitudinal changes and, for longitudinal mammography risk prediction, one aligned recent prior may be more informative than unaligned or multiple prior exams. Uncertainty-based referral has potential to support clinician review but only weakly separates more reliable from less reliable cases in this setting. The code is available at: https://github.com/tpiltner/Temporal-Breast-Cancer-Risk-Prediction.git 

Table of Contents

Acknowledgements 

Chapter 1: Introduction 

Chapter 2: Related Work 

2.1 Single-Exam Risk Prediction Models

2.2 Temporal Mammogram Modeling 

2.3 Spatial Alignment

2.4 Uncertainty Quantification 

2.5 Comparative Literature Review Analyses 

Chapter 3: Dataset

3.1 EMBED Open Dataset

3.2 Choosing Subset of Data

3.2.1 Image Type Filtering 

3.2.2 Standard View Completeness 

3.2.3 Temporal Constraint

3.2.4 Labeling

3.2.5 Final Cohort 

Chapter 4: Approach 

4.1 Pre-Processing

4.2 Image Encoder

4.3 Image Alignment 

4.4 Temporal Information Input

4.5 Cumulative Probability 

4.6 Loss Function

4.7 Training

Chapter 5: Models

5.1 Baseline Model (Current Scan Only)

5.1.1 Model Architecture

5.1.2 Results

5.2 Temporal Model (Current and Most Recent Prior Scan,

Non-Aligned)

5.2.1 Model Architecture

5.2.2 Results

5.3 Temporal Model (Current and Multiple Prior Scans,

Non-Aligned)

5.3.1 Model Architecture

5.3.2 Results

5.4 Temporal Model (Current and Most Recent Prior Scan, Aligned)

5.4.1 Model Architecture

5.4.2 Results

5.5 Temporal Model (Current and Multiple Prior Scans, Aligned)

5.5.1 Model Architecture 

5.5.2 Results 

5.6 Comparative Results 

Chapter 6: Uncertainty Quantification

6.1 Uncertainty Quantification Method

6.2 Referral Policy

6.3 Results

Chapter 7: Conclusions

7.1 Discussion 

7.2 Limitations and Future Research Directions

Bibliography 

About this Honors Thesis

Rights statement

• Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.

School	Emory College
Department	Computer Science
Degree	B.S.
Submission	Honors Thesis
Language	English
Research Field	Computer Science
Keyword	image alignment healthcare medical imaging breast cancer risk prediction machine learning
Committee Chair / Thesis Advisor	Shengpu Tang, Emory University
Committee Members	Nosayba El-Sayed, Emory University Andreas Zufle, Emory University

Last modified

Primary PDF

Thumbnail	Title	Date Uploaded	Actions
	Temporal Mammogram Machine Learning for Breast Cancer Risk Prediction with Uncertainty-Quantified Decision Referrals ()	2026-04-08 10:30:06 -0400	Press to Select an action Download

Supplemental Files

Thumbnail	Title	Date Uploaded	Actions