FROM SCREEN TO DIAGNOSIS: HOW ARTIFICIAL INTELLIGENCE IS REDEFINING MAMMOGRAPHIC INTERPRETATION IN BREAST CANCER SCREENING

Abstract

Objective: To analyze, through an integrative literature review, the efficacy and accuracy of artificial intelligence (AI) applied to mammogram interpretation for breast cancer diagnosis, comparing its results with traditional methods performed by medical specialists. Method: Integrative literature review conducted according to PRISMA 2020 guidelines, with searches performed in PubMed/MEDLINE, Scopus, and the Virtual Health Library (VHL), covering publications from January 2020 to April 2026. Original studies were included — randomized clinical trials, prospective and retrospective cohort studies, and diagnostic accuracy studies — evaluating deep learning-based AI systems applied to mammography, with outcomes of sensitivity, specificity, and/or area under the ROC curve (AUC), compared to radiologist performance. Methodological quality was assessed using QUADAS-2, RoB 2.0, and ROBINS-I tools. Results: Twenty-three studies were included, totaling more than 2.5 million mammograms evaluated, published between 2022 and 2026 in 14 countries. Four randomized clinical trials demonstrated non-inferiority of the AI + radiologist combination compared to standard double human reading, with a 44 to 87% reduction in radiologist workload. AUC values ranged from 0.86 to 0.97, with sensitivity superior or equivalent to human reading in most studies. AI identified up to 54% of false-negative cancers missed by radiologists, with a median diagnostic anticipation of 272 days. AI assistance significantly improved the performance of less experienced radiologists in all studies that assessed this variable. Risk of bias assessment classified 78% of studies as low risk and 22% as uncertain risk; none were classified as high risk. Conclusion: Artificial intelligence presents diagnostic performance non-inferior to conventional human reading in screening and diagnostic mammography, with the potential to reduce radiologist workload, detect false-negative cancers, shorten time to diagnosis, and equalize performance among radiologists of different experience levels. Implementation of AI as a second reader or decision-support tool, with continuous performance monitoring and structured arbitration for discordant cases, is the model best supported by current evidence and can be considered safe for adoption in population-based mammographic screening programs.