Discrepancies in breast cancer screening and hypothesis of artificial intelligence support

Abstract In Italy,the National Breast Screening Programme offers mammography as the 1st level examination, blindly assessed by two radiologists.In the context of continuous quality control,discrepancies in the assessment of mammograms between the 1st and 2nd reader have been studied,focusing on the most significant ones.The aims of the present study are:to evaluate the outcome of women sent to the 2nd level of review;to evaluate the possibility of using artificial intelligence (AI) software to assist radiologists in case of discrepancies. In the Local Health Authority Roma2,radiologists use a standardised classification that rates radiographs on a five-level scale from R1 to R5:R1=no abnormalities,R2=benign findings,R3=equivocal findings,R4=suspected cancer,R5=strongly suspected cancer.Data were extracted from the mammograms performed in 2023 that showed discrepancies (R1/R2 vs. R4/R5).2nd level mammography≥R4 and 2nd level histology≥B4 were considered positive. For the second aim,”IA Lunit INSIGHT” software was used to create specific target groups using the following cut-offs:G1<10%, G2 11-29%; G3 30-59%; G4 60-85%; G5 86-100%(groups ≥G3 were considered positive). Of the 36,339 mammograms,524(1.5%) were discordant:347(66.2%) were considered suspicious by the second radiologist.49(9.4%) were confirmed positive by Level II mammography and 23(4.4%) reported histology≥B4. The AI red a random sample of 247 mammograms and classified 34 women as positive (1 G5, 6 G4, 9 G3), of whom 11 (4.5%) were confirmed positive by Level II mammography and 6 (2.4%) were histologically positive. Of the negative women, 2 had histology≥B4. Discrepancies may be one of the factors leading to inappropriate referral to second level investigation. In such cases, the use of AI may improve the appropriateness of further investigation, although this hypothesis needs to be tested in larger numbers, with additional comparisons with experienced radiologists and consideration of retraining. Key messages • Continuous monitoring in cancer screening programs are vital to assure a sustainability of the process. • Artificial Intelligence software are promising in aiding health care professionals in case of discrepancies.