Session 1: Artificial Intelligence Performance in Breast Cancer Detection: Association of Artificial Intelligence Case Scores and Cancer Characteristics

Purpose: Commercially available Artificial intelligence (AI) software is designed to augment cancer detection on mammograms. The goal of our study is to assess whether a commercially available product that has been approved by the FDA assigns different case scores to breast cancer with different imaging features and pathology.

Materials and Methods: We conducted a retrospective chart review of screening-detected breast cancer cases at a large community practice. We reviewed cases from the year 2021 and used linear regression to determine associations between the overall score given to a case by the AI algorithm (0 being the lowest and 100 being the most severe) and various clinical factors. STATA was utilized for statistical analysis.

Results: Our study included 735 patients with screening-detected breast cancer, with an average age of 65 ± 12. Of the patients, 73% were white, 11% were black, 9% were Hispanic, 3% were Asian, and 5% were of other or unknown race. The average case score was 79±22. We found that the average case scores were 91±13 for mass with calcifications on mammograms, 88±21 for calcifications alone, 76±22 for masses without calcifications, 59±25 for architectural distortion without masses, and 41±33 for lymph node metastasis only. Furthermore, the average case scores were 84±22 for ductal carcinoma in situ, 77±22 for invasive ductal carcinoma, 76±21 for invasive lobular carcinoma, and 62±25 for other types of invasive carcinoma. Interestingly, older patients were more likely to have higher case scores (β-coefficient=0.17, 95% CI 0.03-0.31, P=0.019).

Conclusion: Our findings suggest that AI assigns different case scores to breast cancer with different imaging presentation and pathology. Our results demonstrate that the AI algorithm tends to assign higher case scores to cases with certain characteristics, such as the presence of masses with calcifications or calcifications alone, suggesting that the algorithm may be more sensitive to these imaging features. Conversely, cases with architectural distortion without masses or lymph node metastasis only received lower average case scores, indicating potential challenges in the algorithm's ability to detect these specific types of breast cancer.

Clinical Relevance Statement: Increased awareness of findings and cancer types associated with different AI case scores can lead to improved breast cancer detection. Future research should focus on enhancing the algorithm's performance across breast cancer subtypes to ensure its effectiveness in clinical practice.