Deep learning algorithms for tumor detection in screening mammography

Abstract

Population-wide mammography screening was fully implemented in Sweden in 1997. The implementation has helped to identify breast cancer at earlier stages and thereby lowered mortality by 30-40%. However, it still has its limitations, many studies have shown a discrepancy between radiologist when assessing mammographic examinations. Additionally, women with very dense breasts have a lower mammographic sensitivity and cancers are easily missed. There is also a shortage on breast radiologists and the workload is increasing due to more women being screened. These challenges could be addressed with the help of artificial intelligence systems. The artificial intelligence system can serve both as an assistant to replace one radiologist in a double-reading setting and as a tool to triage women with a high risk of breast cancer for additional screening using other modalities.

In this thesis we used data from two cohorts: the cohort of screen aged women (CSAW) and the ScreenTrust MRI cohort. The primary objectives were to establish performance benchmarks based on radiologists recorded assessments (study I), compare the diagnostic performance of various AI CAD systems (study II), investigate differences and similarities in false assessments between AI CAD and radiologists (study III), and evaluate the potential of artificial intelligence in triaging women for complementary MRI screening (study IV). The data for studies I-III were obtained from CSAW, while the data for study IV were obtained from the MRI ScreenTrust cohort. CSAW is a collection of data from Stockholm County between the years of 2008 and 2015.

Study I was a retrospective multicenter cohort study that examined radiologist performance benchmarks in screening mammography. Operating performance was assessed in terms of abnormal interpretation rate, false negative rate, sensitivity, and specificity. Measures were determined for each quartile of radiologists classified according to performance, and performance was evaluated overall and by different tumor characteristics. The study included a total of 418,041 women and 1,186,045 digital mammograms, and involved 110 radiologists, of which 24 were defined as high-volume readers. Our analysis revealed significant differences in performance between high-volume readers, as well as a variability in sensitivity based on tumor characteristics. This study was presented during the 2019 annual meeting of the Radiological Society of North America, and was awarded the Trainee research prize that same year.

Study II was a retrospective case-control study that evaluated the performance of three commercial algorithms. We performed an external evaluation of these algorithms and compared the retrospective mammography assessments of radiologists with those of the algorithms. Operating performance was determined in terms of abnormal interpretation rate, false negative rate, sensitivity, specificity and the AUC. The study included 8,805 women, of whom 740 women had cancer, and a random sample of 8,066 healthy controls. There were 25 radiologists involved. For a binary decision, the cutpoint was defined by the mean specificity of the original first-reader radiologists (96.6%). Our findings showed that one AI algorithm outperformed the other AI algorithm and the original first-reader radiologists. This study was presented during the 2020 annual meeting of the European Society of Radiology.

Study III was a retrospective case-control study that evaluated the differences and similarities in false assessments between an artificial intelligence system and a human reader in screening mammography. In this study we included 714 screening examinations for women diagnosed with breast cancer and 8,003 randomly selected healthy controls. The abnormality threshold was predefined from study II. We examined how false positive and false negative assessments by AI CAD and the first radiologist, were associated with breast density, age and tumor characteristics. Our findings showed that AI makes fewer false negative assessments than radiologists. Combining AI with a radiologist resulted in the most pronounced decrease in false negative assessments for high-density women and women over the age of 55. This study was presented at the 2021 annual meeting of the Radiological Society of North America.

Study IV is a randomized clinical trial that aims to investigate the effect of applying deep learning methods to select women for MRI-based breast cancer screening. The study examines how effectively AI can identify women who should be offered a complementary MRI screening based on their likelihood of having cancer that is not visible on regular mammography. The results reported in this thesis are preliminary and based on examinations from April 1, 2021 to December 31, 2022. During the indicated time period, 481 MRI examinations have been completed, and 28 cancers have been detected, yielding a cancer detection rate of 58.2 per 1,000 examinations. Although, the trial is still ongoing, the inter-rim results suggest that using AI-based selection for supplemental MRI screening can lead to a higher rate of cancer detection than that reported for density-based selection methods.

In conclusion, we have shown that the use of AI for breast cancer detection can increase precision and efficiency in mammography screening.