A recent Lancet Regional Health study assesses the performance of an artificial intelligence (AI)-based risk model for breast cancer screening in Europe.

Study: European validation of an image-derived AI-based short-term risk model for individualized breast cancer screening—a nested case-control study. Image Credit: Gagliardiphotography / Shutterstock.com

Background

Regular mammography screening has reduced deaths due to breast cancer in women. Even after biennial screening for breast cancer, about 25% of breast cancers are diagnosed. In these cases, some women might have tested negative in one mammographic screening but could have been diagnosed with breast cancer before attending their next screening appointment.

Between 25-40% of women are diagnosed with breast cancer at stage two or higher. Thus, it is important to determine whether the tumor was detected during the regular mammographic screening, as it is a robust prognostic marker of breast cancer-related mortality.

Previous studies have proposed the addition of other risk assessment measures to improve the screening process and ultimately prevent the risk of interval cancer before the next screen. This strategy could also reduce the incidence of late-stage breast cancer in the next screen. In the United States, women who have dense breasts or are at a high risk due to familial risk factors, undergo additional examinations.

The current breast cancer screening programs conducted in Europe do not have any guidelines that indicate the performance of additional examinations for women at a higher risk of breast cancer. However, several clinical risk assessment tools have been developed based on family history and lifestyle factors to improve screening outcomes.

Although a new image-based risk model has shown considerable potential in identifying women at a higher risk of breast cancer, this model requires additional external validation to assess its clinical feasibility.

About the study

The current study assessed a previously developed image-derived AI-based risk model for breast cancer that was designed to identify the risk of breast cancer in the short term. More specifically, this model has been used to identify women who developed cancer in the interval between two mammography screenings in two years after a negative screen.

The overall risk classification and discriminatory performance of the ProFound AI Risk model were assessed. This AI-based model was previously developed using a screening Swedish cohort.

The current study used four screening populations comprising women between 45 and 69 years of age who underwent mammographic screening. From this screening population, two cohorts were designed in Germany and one each from Italy and Spain.

Some of the key eligibility criteria included the incidence of breast cancer with a digital mammogram at baseline. These women were diagnosed before or at the next screening program. 

The study excluded women with a family history of breast cancer. A nested case-control study for each population was performed. Control groups for each screening population were randomly designed from the underlying screening cohort.

Study findings

The validation study included a total of 739 breast cancer patients and 7,812 controls. The cancer outcome was assessed at the second screen, during which women were randomly assigned to have digital mammography or were subjected to digital breast tomosynthesis (DBT). The AI-based risk model used these mammographs to predict women who were at risk of breast cancer in two years.

As compared to the original assessment of the AI-based risk model for breast cancer screening that used a Swedish cohort, a small variability of discriminatory performances across populations of different European countries was observed. However, the model exhibited similar discrimination to that of the previous report. Women with dense and non-dense breasts exhibited similar risk stratification performance.

Advanced-stage breast cancer was most likely to be diagnosed in high-risk women as compared to those at a moderate risk of developing breast cancer. The current study indicated that an image-based AI-risk model could be affected by ethnic differences and screening frequencies.

Women with non-dense breasts were found to be at a greater risk of developing more aggressive interval cancers. In contrast, women with dense breasts could have their tumor masked by dense tissue, which increases the possibility of developing interval cancer and late-stage breast cancer.

Radiologists experience significant challenges related to the masking of tumors by dense tissues. Therefore, high-risk women with dense breasts could positively benefit from more sensitive examinations following a negative screening. Nevertheless, a shorter screening interval is preferable for high-risk women with non-dense breasts due to the increased risk of a fast-growing tumor. 

Conclusions

The current study provided insights into the importance of conducting additional tests beyond mammographic density to identify women who are at a higher risk of breast cancer, which would positively improve screening outcomes. A combination of density and risk assessment approaches could be more effective in population-based screening programs for breast cancer.