Purpose: Artificial intelligence (AI) is changing the practice of medicine, including early detection of breast cancer with computer-assisted detection (CAD). Early studies suggest AI-CAD can improve cancer detection rates (CDR) and recall rates. However, the majority of these studies used cancer-enriched datasets in non-clinical environments or were performed outside the US. This study assesses the impact of AI-CAD in clinical use in the US.
Materials and Methods: We performed a retrospective analysis of screening mammograms with digital breast tomosynthesis in our practice including an academic institution, community breast center, mobile mammography unit, and two standalone screening centers. Data were compared during a pre-AI-CAD period during which standard Hologic R2 CAD was used (7/1/2022-12/31/2022) and a post-AI-CAD group (2/1/2023-7/31/2023). An FDA-approved AI-CAD software was implemented in January 2023 and data from this transition month was excluded. Interpretations were made by a group of 11 academic breast radiologists with experience of 2 to 38 years.
Data from our mammography reporting system (MagView) were used to calculate CDR and recall rates and chi-square test was used for comparison. Case scores for each mammogram ranged from 0-100 (0 being the lowest likelihood of cancer). Distribution of case scores in the post-AI-CAD group and case scores of patients with cancer were exported from our PACS (Sectra).
Results: In the pre-AI-CAD period, 15684 screening mammograms were performed; CDR was 7.59 per 1000 (119/15684) and recall rate was 10.7% (1677/15684). In the post-AI-CAD period, CDR increased to 7.93 per 1000 (122/15371) and recall rate increased to 11.2% (1722/15371). CDR increased by 4.5% but difference was not statistically significant (p=0.72 for CDR, 0.15 for recall rate).
Cancers were more frequent in screening mammograms with higher case scores (Figure 1) but were found in patients in each decile (Figures 2, 3, and 4). Invasive cancers comprised 74% of AI-CAD detected cancers as compared to 70% without AI-CAD (Figure 5).
Conclusion: Although there is potential for AI-CAD improve CDR and recall rate, analysis of early clinical implementation of AI-CAD did not demonstrate significant difference in either metric in our academic practice. A significant proportion of cancers were found in patients with low case scores and independent review by radiologist is still necessary.
Clinical Relevance Statement: Understanding the impact of AI-CAD systems in clinical use will inform decisions on implementation of AI in cancer screening programs.
