Abstract PO2-29-04: An AI-based safeguard process to reduce aggressive missed cancers in dense breasts at screening mammography

Abstract

Abstract Background: Dense breast tissue makes cancerous lesions difficult to identify on mammograms, resulting in many cancers missed at screening in patients with dense breasts. These missed cancers are then only diagnosed once they have grown to the point they can be more easily distinguished from the surrounding dense tissue. Dense breast tissue is more common in younger women, for whom early detection and diagnosis could lead to fewer years of life lost. This challenge motivated the FDA’s decision this year to require breast density to be included in patient reports, but notification does not solve the problem that these cancers are still often missed. Missing aggressive, fast growing cancers is particularly worrisome given they are more likely to progress to a stage where they are difficult to treat by the time they are detected in the future. Methods: A prospective study was conducted at one group practice of 6 radiologists over the course of 12 months to test a safeguard process to detect missed cancers so they could be treated more readily. AI software (Saige, DeepHealth, Inc.) was used to flag the most suspicious screening mammography exams that had not been recalled for work-up by an initial interpreting radiologist. An expert breast imager performed a safeguard review of each exam flagged. If the expert decided a recall was warranted, she would consult with the initial interpreting radiologist to decide if the patient should be recalled for further (diagnostic) imaging. Data was collected for each breast cancer diagnosed during this period. Patient data included breast density, age, and race. Cancer follow up data included pathology classification, hormone status, grade, staging, and Ki67 score. Aggressive cancers were identified as those with high grade (Grade 3), high Ki67 ( >=20%) or clinical stage of IIA or greater. A comparison was performed between cancers detected by the original interpreting radiologist and those detected through the safeguard process. Results: During the study period, 40,532 mammograms were obtained and 2,296 were flagged for safeguard review. The safeguard review led to 130 additional patient recalls and 41 additional cancer diagnoses. In women with dense breast tissue, 103 cancers were detected by routine interpretation and an additional 18 cancers (17.5%) detected through the safeguard process. In women without dense breast tissue, 116 cancers were detected by routine interpretation and 23 additional cancers were detected through the safeguard process. The cancers detected through the safeguard process included a larger percent of aggressive cancers (34.1% vs 30.6%, p=0.66) and more than double the percent of triple-negative cancers (9.8% vs 4.1%, p=0.13), though these differences were not statistically significant. The cancers detected through the safeguard process were from patients with similar demographics (age and race) to those whose cancers were detected by the interpreting radiologist. Conclusion: The AI-driven safeguard process results in a significant increase in breast cancers that would have been missed in patients with dense breast tissue. Many of the additional cancers were aggressive with almost 10% being triple-negative cancers. Without the safeguard process those aggressive cancers would not likely have been diagnosed until the next screening cycle in 1-2 years when treatment would likely have been less effective and outcomes poorer. Citation Format: Bryan Haslam, Jiye Kim, A. Gregory Soresen. An AI-based safeguard process to reduce aggressive missed cancers in dense breasts at screening mammography [abstract]. In: Proceedings of the 2023 San Antonio Breast Cancer Symposium; 2023 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2024;84(9 Suppl):Abstract nr PO2-29-04.