AI-based prevention of interval cancers in a national mammography screening program.

PURPOSE: To demonstrate that artificial intelligence (AI) can detect and correctly localise retrospectively visible cancers that were missed and diagnosed as interval cancers (false negative (FN) and minimal signs (MS) interval cancers), and to characterise AI performance on non-visible occult and true interval cancers. METHOD: Prior screening mammograms from N = 2,396 women diagnosed with interval breast cancer between March 2006 and May 2018 in north-western Germany were analysed with an AI system, producing a model score for all studies. All included studies previously underwent independent radiological review at a mammography reference centre to confirm interval cancer classification. Model score distributions were visualised with histograms. We computed the proportion and accompanying 95% confidence intervals (CI) of retrospectively visible and true interval cancers detected and correctly localised by AI at different operating points representing recall rates < 3%. Clinicopathological characteristics of retrospectively visible cancers detected by AI and not were compared using the Chi-squared test and binary logistic regression. RESULTS: Following radiological review, 15.6% of the interval cancer cases were categorised as FN, 19.5% MS, 11.4% occult, and 53.4% true interval cancers. At an operating point of 99.0% specificity, AI could detect and correctly localise 27.5% (95% CI: 23.3-32.3%), and 12.2% (95% CI: 9.5-15.5%) of the FN and MS cases on the prior mammogram, respectively. 228 of these retrospectively visible cases were advanced/metastatic at diagnosis; 21.1% (95% CI: 16.3-26.8%) were found by AI on the screening mammogram. Increased likelihood of detection of retrospectively visible cancers with AI was observed for lower-grade carcinomas and those with involved lymph nodes at diagnosis. Among true interval cancers, AI could detect and correctly localise in the screening mammogram where subsequent malignancies would appear in 2.8% (95% CI: 2.0-3.9%) of cases. CONCLUSIONS: AI can support radiologists by detecting a greater number of carcinomas, subsequently decreasing the interval cancer rate and the number of advanced and metastatic cancers.

[Time trend in programme sensitivity within the German mammography screening programme in North Rhine-Westphalia and Lower Saxony]. / Zeitliche Entwicklung der Programmsensitivität des deutschen Mammographie-Screening-Programms in Nordrhein-Westfalen und Niedersachsen.

BACKGROUND: The programme sensitivity is a performance indicator for evaluating the quality of the mammography screening programme (MSP). OBJECTIVES: We analysed the development of the programme sensitivity over time in two federal states of Germany, North Rhine-Westphalia (NRW) and Lower Saxony (NDS). MATERIALS AND METHODS: Data from 2,717,801 (NRW) and 1,197,660 (NDS) screening examinations between 2006 and 2011 were linked with data of the State Cancer Registry NRW and the Epidemiological Cancer Registry NDS, respectively. Breast cancers (invasive and in situ) were either detected at screening or diagnosed within the 24-month interval after an inconspicuous screening result outside the programme. The crude and age-standardized programme sensitivity was calculated per calendar year. The German mammography screening office provided aggregated recall rates. RESULTS: The age-standardized programme sensitivity increased markedly for initial screening examinations from 2006 to 2011 from 75.0% (95% CI: 72.1-77.9) to 80.5% (95% CI: 78.5-82.5) in NRW, and from 74.9% (95% CI: 71.4-78.5) to 84.7% (95% CI: 81.1-88.3) in NDS. Concurrently, recall rates increased as well. For subsequent screening examinations, the programme sensitivity increased from 2008 to 2011 from 68.1% (95% CI: 63.1-73.1) to 71.9% (95% CI: 70.2-73.6) in NRW, and from 69.8% (95% CI: 64.2-75.4) to 74.9% (95% CI: 72.3-77.5) in NDS, whereas the recall rates remained relatively constant. CONCLUSIONS: In both federal states, the programme sensitivity increased over time. This increase, possibly indicating an improved quality of diagnosis within the MSP as a learning system, is discussed under consideration of the age distribution of screening participants and the recall rates.