Predicting short- to long-term breast cancer risk from longitudinal mammographic screening history

Abstract

Breast cancer (BC) risk assessment aims to enhance individualized screening and prevention strategies. While recent deep learning (DL) models based on mammography have shown promise in short-term risk prediction, they primarily rely on single-time-point (STP) exams, ignoring temporal changes in breast tissue from sequence exams. We present the Multi-Time Point Breast Cancer Risk Model (MTP-BCR), a novel DL approach that integrates traditional risk factors and longitudinal mammography data to capture subtle tissue changes indicative of future malignancy. Using a large in-house dataset with 171,168 mammograms from 9133 women, MTP-BCR achieved superior performance in 10-year risk prediction, with an AUC of 0.80 (95% CI, 0.78-0.82) at the patient level, outperforming STP-based and traditional risk models. External validation on the CSAW-CC dataset confirmed its robustness. Further analysis demonstrates the advantages of the MTP-BCR method in diverse populations. MTP-BCR also excels in risk stratification and offers heatmaps to enhance clinical interpretability.

Fig. 1. Overview of the study design.

A In-house and public datasets collection for model building and validation. We build our model based on the in-house dataset. To further evaluate the model’s performance on two screening cohorts, a recurrence cohort, and a public dataset. B A schematic diagram of a retrospective patient trajectory. The diagram simulates timelines of retrospective patient trajectories for the in-house dataset. Each grey dot represents a mammogram examination, and red dots mean the diagnostic examination with pathology results. C Overview of retrospective patients’ trajectory for the in-house dataset. In the first heatmap, we plot the distribution of examinations in terms of the number of prior reference examinations and the time to last negative follow-up in the no cancer group or the time to diagnosis in the cancer group. In the second, we plot the distribution of examinations in terms of the time interval between the target and the earliest prior reference examinations and the time to last negative follow-up in the no cancer group or the time to diagnosis in the cancer group. The plot shows that the retrospective trajectory distributions of the cancer and no cancer groups are similar. D Schematic of the BC risk prediction by the MTP-BCR model. The target mammogram (the dot with a green circle) is the mammogram for which the risk is calculated. This can be any mammogram in the timeline of a woman. Prior mammograms are those obtained before the target mammogram, and up to five prior mammograms can be inputted (the grey or red dots with blue circles). The time to BC is the time between eventual cancer detection and the target mammogram, which can be up to 10 years. E The architecture of the proposed MTP-BCR model. (Detailed in the Method).

Fig. 2. ROC curves and ablation experiments results of the risk prediction of the MTP-BCR model.

A ROCs for the MTP-BCR method patient- and breast-level prediction on the entire in-house test set. B Ablation experiments results on the entire in-house test set. The first plot shows that our proposed learning strategies can improve the ability of the risk model. The second plot shows that the performance of our risk model improves with more of the prior reference exams. C ROCs for MTP-BCR method patient- and breast-level prediction on the CSAW-CC dataset. D–F ROCs for risk prediction on three in-house cohorts.

Fig. 3. Comparison results of risk prediction.

A The comparison results of age-adjusted AUC (aAUC) based on the different cohorts. Traditional risk model (BCSC) based on risk factors; Baseline-risk factors: SVM model based on risk factors; Baseline-recurrence SVM: a traditional machine learning (SVM)-based recurrence risk prediction model, leveraging risk factors and prognostic factors. STP-baseline: Single-time point (STP) image-only based baseline DL methods; STP-detection: STP-based DL detection method; STP-transformer: STP-based DL risk prediction method, which leverages the transformer to fuse the representation of each view. B Kaplan–Meier (KM) curves for 10-year risk stratifying by different methods on primary screening cohort 2. Groups are divided by the predicted risk scores of each model, including (1) 0 to 10th percentile, (2) 10th to 50th percentile, (3) 50th to 90th percentile, and (4) 90th and up. The summary table for each percentile range was provided, detailing the number of women, the number of cancers, the percent of cancers accounted for women of each group, and the percent of group cancers accounted for all cancers.

Fig. 4. Cumulative risk at multiple time points on different sub-groups.

A–D AUC performance of the MTP-BCR model across different subgroups, based on (A) age, (B) breast density, (C) molecular subtypes, and (D) receptor subtypes of future tumors.

Fig. 5. An example of a class activation map (CAM) visualization.

The longitudinal craniocaudal (CC) and mediolateral oblique (MLO) mammograms were acquired from a patient who participated in ten consecutive BC screening from 2005 to 2015, culminating in a BC diagnosis during the last screening (invasive ductal and lobular carcinoma located at C50.4, exhibiting positive expression of estrogen receptor (ER+), progesterone receptor (PR+), and human epidermal growth factor receptor 2 (Her2Neu+)). The closer to red, the more relevant the pixel is to the risk prediction.