Deep Learning for Breast Cancer Risk Prediction: Application to a Large Representative UK Screening Cohort

Abstract

Purpose To develop an artificial intelligence (AI) deep learning tool capable of predicting future breast cancer risk from a current negative screening mammographic examination and to evaluate the model on data from the UK National Health Service Breast Screening Program. Materials and Methods The OPTIMAM Mammography Imaging Database contains screening data, including mammograms and information on interval cancers, for more than 300 000 female patients who attended screening at three different sites in the United Kingdom from 2012 onward. Cancer-free screening examinations from women aged 50-70 years were performed and classified as risk-positive or risk-negative based on the occurrence of cancer within 3 years of the original examination. Examinations with confirmed cancer and images containing implants were excluded. From the resulting 5264 risk-positive and 191 488 risk-negative examinations, training (n = 89 285), validation (n = 2106), and test (n = 39 351) datasets were produced for model development and evaluation. The AI model was trained to predict future cancer occurrence based on screening mammograms and patient age. Performance was evaluated on the test dataset using the area under the receiver operating characteristic curve (AUC) and compared across subpopulations to assess potential biases. Interpretability of the model was explored, including with saliency maps. Results On the hold-out test set, the AI model achieved an overall AUC of 0.70 (95% CI: 0.69, 0.72). There was no evidence of a difference in performance across the three sites, between patient ethnicities, or across age groups. Visualization of saliency maps and sample images provided insights into the mammographic features associated with AI-predicted cancer risk. Conclusion The developed AI tool showed good performance on a multisite, United Kingdom-specific dataset. Keywords: Deep Learning, Artificial Intelligence, Breast Cancer, Screening, Risk Prediction Supplemental material is available for this article. ©RSNA, 2024.

Keywords: Artificial Intelligence; Breast Cancer; Deep Learning; Risk Prediction; Screening.

Graphical abstract

Figure 1:

Patient-level data selection flowchart. Reported numbers (n) refer to women. From the OPTIMAM database, suitable examinations were identified as those that did not show malignant findings and contained sufficient follow-up information to be classified as risk-positive or risk-negative. After images with implants were removed, the remaining women were then split in a stratified manner into training, validation, and test datasets for model training and evaluation. For computational reasons, numbers of women in the training and validation risk-negative groups were reduced via random sampling.

Figure 2:

Receiver operating characteristic (ROC) performance of artificial intelligence risk model on the hold-out test dataset. The predictive performance of patient age alone is also shown for comparison. AUC = area under the ROC curve, FPR = false-positive rate, TPR = true-positive rate.

Figure 3:

Receiver operating characteristic (ROC) performance of artificial intelligence risk model with respect to (A) screening site, (B) patient age (in years) at screening, (C) patient ethnic background, and (D) mammography unit model. AUC = area under the ROC curve, FPR = false-positive rate, TPR = true-positive rate.

Figure 4:

Receiver operating characteristic (ROC) performance of the presented artificial intelligence risk model with respect to subsequent cancer type. In this case, the ROC curve was calculated using the same negative cases for all cancer types, whereas for Figure 3, the negative cases were also split by the demographic factor under consideration. AUC = area under the ROC curve, FPR = false-positive rate, TPR = true-positive rate.

Figure 5:

Example images from mammography examinations classified by the developed artificial intelligence tool as risk-positive and risk-negative, including correct and incorrect classifications.

Figure 6:

Saliency map demonstration for two example images. The normal baseline craniocaudal (A) and mediolateral oblique (D) mammograms are shown along-side the same images with saliency map overlays (B, E), and the future mammograms containing screen-detected cancers (C, F, with lesions shown in red boxes). White dashed lines highlight the zoomed regions.