Automated measurements for personalized breast cancer screening
Quantitative image information combined with personal risk factors can provide breast screening that is better tailored to the individual.
In the UK, the National Health Service Breast Screening Programme offers all women mammography every three years as part of a breast cancer screening initiative. The only exception is for individuals at moderate and high risk because of their family history, who are eligible for enhanced surveillance using mammography and MRI.1 At the University of Manchester, we have been conducting a study of women undergoing routine breast screening: Predicting Risk Of Cancer At Screening (PROCAS). The study explores taking quantitative measurements from mammograms and using these to inform women of their risk of developing cancer, while also ensuring that risk-reducing strategies are made available where appropriate. In this way, PROCAS could inform a more personalized breast screening plan. However, during the study we collected data that suggests that, based on non-mammographic risk factors alone, in more than 1% of women the 10-year risk of developing breast cancer is greater than 8%, with a further 8.2% of women having a moderately increased risk of 5–8%.2 Therefore, we are now also investigating the feasibility of adapting screening to the individual using both personal risk factors and quantitative measurements taken from screening mammograms.
Breast density describes the proportion of glandular tissue in the breast, and its relationship to the risk of developing breast cancer is well established.3, 4 The PROCAS study collected mammographic density data from more than 50,000 women. Two expert readers visually assessed the data, and we used software systems Volpara and Quantra for automated measurements. Incorporation of visually assessed density adds useful information to breast cancer risk models,5 substantially increasing the proportion of women accurately assigned to high- and low-risk groups. However, visual assessment is subject to both inter- and intra-observer variability.6 While it is possible to correct values for bias across readers,7 it would be preferable to move from subjective, area-based assessment to objective, automated volumetric measurement.8 One difficulty is a lack of agreement between automated methods, which can lead to a woman being assessed as high density by one algorithm, and low by another. Figure 1 illustrates the lack of agreement between the three methods for women with high density. Some of the discrepancies can be explained by image features and patient positioning,9 but the differences pose a real challenge when accurate measurements of density are required to make clinical decisions.
The relationship between mammographic density and non-mammographic risk factors for breast cancer is an active research area. In the PROCAS study, we collected risk factor data by questionnaire, where women self-reported factors such as height, weight, family history, hormone replacement therapy (HRT) use, and menopausal status. While it is possible to identify obvious errors, the accuracy of such data remains open to question. Weight is of particular interest, because mammographic density is often expressed as a relationship between the glandular and fatty components of the breast, so density measurements are often adjusted to take body mass index (BMI) into account when assessing cancer risk. As an alternative approach, automated breast density software can measure the volume of fat in the breast, which is significantly correlated with weight, but this method is unlikely to be an adequate surrogate for all women.10
In our study, visually assessed density showed the strongest relationship with risk of cancer. However, it is likely that, while readers were asked to quantify density, they also took into account higher-level features, such as the pattern and location of dense areas. To date, we have shown that density is increased in the area of the mammogram in which cancer will develop in the future.11 When developing and evaluating novel density and texture methods, the availability of high-quality data is of paramount importance. Any change in calibration of mammography systems over time can affect performance measurements, unless controls are carefully matched to cancer cases. Differences in age, BMI, parity, HRT use, and menopausal status can have similar effects. Furthermore, full-field digital mammography has only recently been introduced for screening. With the three-year screening interval in the UK, and the lack of systematic collection of unprocessed digital mammogram data, longitudinal data for previous screening mammograms of women who subsequently develop breast cancer is only just becoming available in sufficient quantity to enable evaluation of risk prediction. Prior to this, research focused on analyzing images of the breast opposite that in which cancer has been detected, using cross-validation to provide a robust and efficient way of exploiting the available data.12
For fully personalized screening, the process of allocating women to different screening regimes is likely to include the risk of cancers being obscured by dense breast tissue (masking), as well as individual risk of developing cancer computed from an established risk model incorporating density and texture measurements extracted from the mammogram. For example, it would not be desirable to recommend an increased frequency of mammographic screening for women in whom mammography is less effective.
In summary, breast density is an important risk factor for breast cancer and adds useful information to breast cancer risk prediction models, although there is a lack of agreement between methods of obtaining the data, which requires further investigation. It is likely that readers also take higher-level features into account when visually assessing density. This may be captured by automated breast texture measurements.
Our ongoing work includes investigating automated measures of mammographic texture13 (as well as breast density), using longitudinal breast cancer screening data with a view to further improving risk prediction models, and exploring how personalized breast screening could work in practice.
The research described was funded by the National Institute for Health Research, Breast Cancer Now, Genesis, and the European Union. All of the work is collaborative, and we wish to acknowledge the support and contribution of colleagues at the University of Manchester, Queen Mary University of London, and the Nightingale Breast Centre and Genesis Prevention Centre, Manchester.
Susan Astley is an imaging scientist and a member of a multidisciplinary team investigating imaging applications in breast cancer detection, diagnosis, and treatment.
Elaine Harkness is an epidemiologist working with a multidisciplinary team investigating the relationship between breast density and the risk of developing breast cancer.