Detection of contralateral breast cancer in patients with primary breast cancer through AI-assisted mammography.

e13645 Background: Survivors of breast cancer are at an elevated risk for developing contralateral breast cancer (CBC), a risk that surpasses the incidence of initial primary breast cancer within the general population. Consequently, annual mammography screenings are advocated to monitor for ipsilateral recurrence and the emergence of CBC. However, the sensitivity of surveillance mammography exhibits considerable variation, ranging between 45% and 90%, with notably lower detection rates in younger individuals and those with dense breast tissue. This study explores the efficacy of an artificial intelligence (AI) detection algorithm in mammography for the early identification of CBC prior to the official diagnosis. Methods: A retrospective analysis was conducted on patients treated for primary breast cancer from 2000 to 2021 at Seoul National University Hospital. This involved reviewing serial mammograms post-surgery and adjuvant therapy from 460 patients who subsequently developed CBC (CBC group) and 469 patients who did not (control group). These mammograms were evaluated using the Lunit INSIGHT mammogram AI system, which provides a pixel-level abnormality score between 0 and 100, with higher scores indicating increased abnormality. Scores exceeding 10 were considered indicative of CBC, reflecting a 90% sensitivity based on the AI software's tuning dataset. The diagnostic accuracy of AI-assisted mammography was then juxtaposed with that of specialist radiologists' evaluations. Results: At the point of CBC diagnosis, the median AI abnormality score was markedly higher in the CBC group (22.9 [IQR, 2.1-81.1]) compared to the control group's final follow-up mammogram (0.4 [IQR, 0.1-1.5], p<0.001). The AI system's sensitivity for CBC detection paralleled that of radiologist assessments (59.6% for AI vs. 56.5% for radiologists out of 460 detected cancers), with the AI method achieving an AUROC of 0.836 and a specificity of 91.0%. Notably, AI detected 84 cases (30.7%) at least six months and 59 cases (21.5%) more than a year before the official CBC diagnosis. Conclusions: The application of AI in mammography screening has demonstrated significant potential for the early detection of CBC, identifying cases up to a year before formal diagnosis. This underscores the value of integrating AI tools in mammography to enhance early detection rates of CBC among primary breast cancer patients.