2024 ARRS ANNUAL MEETING - ABSTRACTS

RETURN TO ABSTRACT LISTING


E4877. Artificial intelligence for Detection of Contralateral Second-in-Breast Cancer in Patients Treated With Unilateral Mastectomy
Authors
  1. Su Min Ha; Seoul National University Hospital
  2. Jung Min Chang; Seoul National University Hospital
Objective:
This study aims to assess the outcome of postoperative mammographic screening with artificial intelligence (AI) software for detection of contralateral second-in-breast cancer in patients treated with unilateral mastectomy.

Materials and Methods:
A retrospective search was conducted on the database of academic medical center for 833 patients with breast cancer treated with unilateral mastectomy between January 2011 and December 2014 and who had undergone postoperative surveillance mammography. Commercially available AI software (Lunit INSIGHT MMG, Ver. 1.1.7.1) was used. For 29 patients with second-in-breast cancers, AI software was applied to two mammographic views (craniocaudal and mediolateral oblique) of the remaining unilateral breast at the time of diagnosis. For 804 patients without recurrence, AI software was applied to incident postoperative mammography. Mammography with AI output results were reviewed, defining maximum pixel-level abnormality score = 10 in at least one view at cancer site as test positive. The cancer detection rate (CDR) per 1000 screening examination, sensitivity, specificity, PPV, NPV, and recall rates were compared.

Results:
Of 29 second-in-breast cancers, 20 were invasive and nine were ductal carcinoma in situ; median invasive cancer size was 0.9 cm (range, 0.1–3.7cm). At the time of diagnosis, mammography alone detected 18 recurrences. AI software alone detected 19 recurrences. Mammography with AI detected 21 recurrences (14 invasive and seven ductal carcinoma in situ). Mammography alone showed CDR of 21 (19/833; 95% CI: 13–34%), sensitivity of 62% (18/29; 95% CI: 43–77%), specificity of 99% (796/804; 95% CI: 98–99%), PPV of 69% (18/26; 95% CI: 49–83%), NPV of 98% (796/807; 95% CI: 97–99%), and recall rate of 3% (26/833; 95% CI: 2–4%). AI software alone showed CDR of 23 (18/833; 95% CI: 14–35%), sensitivity of 65% (19/29; 95% CI: 46–80%), specificity of 92% (741/804; 95% CI: 90–93%), PPV of 23% (19/82; 95% CI: 15–33%), NPV of 98% (741/751; 95% CI: 97–99%), and recall rate of 9% (82/833; 95% CI: 7–12%). Mammography with AI showed CDR of 25 (21/833; 95% CI: 16–38%), sensitivity of 72% (21/29; 95% CI: 53–85%), specificity of 91% (736/804; 95% CI: 89–93%), PPV of 23% (21/89; 95% CI: 15–33%), NPV of 98% (736/744; 95% CI: 97–99%), and recall rate of 9% (78/833; 95% CI: 7–11%). Compared to mammography alone, addition of AI software to mammography showed higher CDR and sensitivity with marginal significance (p =. 083, and .069, respectively), but with lower specificity and higher recall rate (p < .001).

Conclusion:
In patients treated with unilateral mastectomy, AI software assists in detection of second-in-breast cancers in women with unilateral postoperative mammography. However, due to the lack of reference mammography image for comparison, there is an increase in false positive results. In the future, AI software performance should be further refined to address the absence of contralateral reference, thereby reducing the number of recalls.