Deep learning algorithms for tumor detection in screening mammography

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 ...