Artificial Intelligence for Image-Based Breast Cancer Risk Prediction Using Attention.

Accurate prediction of individual breast cancer risk paves the way for personalised prevention and early detection. The incorporation of genetic information and breast density has been shown to improve predictions for existing models, but detailed image-based features are yet to be included despite correlating with risk. Complex information can be extracted from mammograms using deep-learning algorithms, however, this is a challenging area of research, partly due to the lack of data within the field, and partly due to the computational burden. We propose an attention-based Multiple Instance Learning (MIL) model that can make accurate, short-term risk predictions from mammograms taken prior to the detection of cancer at full resolution. Current screen-detected cancers are mixed in with priors during model development to promote the detection of features associated with risk specifically and features associated with cancer formation, in addition to alleviating data scarcity issues. MAI-risk achieves an AUC of 0.747 [0.711, 0.783] in cancer-free screening mammograms of women who went on to develop a screen-detected or interval cancer between 5 and 55 months, outperforming both IBIS (AUC 0.594 [0.557, 0.633]) and VAS (AUC 0.649 [0.614, 0.683]) alone when accounting for established clinical risk factors.

Special Section Guest Editorial: Advances in Breast Imaging.

This guest editorial introduces the special section on Advances in Breast Imaging.

Evaluation of a Stratified National Breast Screening Program in the United Kingdom: An Early Model-Based Cost-Effectiveness Analysis.

OBJECTIVES: To identify the incremental costs and consequences of stratified national breast screening programs (stratified NBSPs) and drivers of relative cost-effectiveness. METHODS: A decision-analytic model (discrete event simulation) was conceptualized to represent four stratified NBSPs (risk 1, risk 2, masking [supplemental screening for women with higher breast density], and masking and risk 1) compared with the current UK NBSP and no screening. The model assumed a lifetime horizon, the health service perspective to identify costs (£, 2015), and measured consequences in quality-adjusted life-years (QALYs). Multiple data sources were used: systematic reviews of effectiveness and utility, published studies reporting costs, and cohort studies embedded in existing NBSPs. Model parameter uncertainty was assessed using probabilistic sensitivity analysis and one-way sensitivity analysis. RESULTS: The base-case analysis, supported by probabilistic sensitivity analysis, suggested that the risk stratified NBSPs (risk 1 and risk-2) were relatively cost-effective when compared with the current UK NBSP, with incremental cost-effectiveness ratios of £16,689 per QALY and £23,924 per QALY, respectively. Stratified NBSP including masking approaches (supplemental screening for women with higher breast density) was not a cost-effective alternative, with incremental cost-effectiveness ratios of £212,947 per QALY (masking) and £75,254 per QALY (risk 1 and masking). When compared with no screening, all stratified NBSPs could be considered cost-effective. Key drivers of cost-effectiveness were discount rate, natural history model parameters, mammographic sensitivity, and biopsy rates for recalled cases. A key assumption was that the risk model used in the stratification process was perfectly calibrated to the population. CONCLUSIONS: This early model-based cost-effectiveness analysis provides indicative evidence for decision makers to understand the key drivers of costs and QALYs for exemplar stratified NBSP.

Detecting and classifying linear structures in mammograms using random forests.

Detecting and classifying curvilinear structure is important in many image interpretation tasks. We focus on the challenging problem of detecting such structure in mammograms and deciding whether it is normal or abnormal. We adopt a discriminative learning approach based on a Dual-Tree Complex Wavelet representation and random forest classification. We present results of a quantitative comparison of our approach with three leading methods from the literature and with learning-based variants of those methods. We show that our new approach gives significantly better results than any of the other methods, achieving an area under the ROC curve A(z) = 0.923 for curvilinear structure detection, and A(z) = 0.761 for distinguishing between normal and abnormal structure (spicules). A detailed analysis suggests that some of the improvement is due to discriminative learning, and some due to the DT-CWT representation, which provides local phase information and good angular resolution.