Artificial intelligence-based tumor size measurement on mammography: agreement with pathology and comparison with human readers' assessments across multiple imaging modalities.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, Kangbuk Samsung Hospital, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea.
- Department of Radiology, Graduate School, The Catholic University of Korea, Seoul, Republic of Korea.
- Department of Radiology, College of Medicine, Seoul St. Mary's Hospital, The Catholic University of Korea, Banpo-Daero 222, Seocho-Gu, Seoul, 06591, Republic of Korea. [email protected].
- Department of Radiology, College of Medicine, Seoul St. Mary's Hospital, The Catholic University of Korea, Banpo-Daero 222, Seocho-Gu, Seoul, 06591, Republic of Korea.
- Division of Biostatistics, Kangbuk Samsung Hospital, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea.
Abstract
To evaluate the agreement between artificial intelligence (AI)-based tumor size measurements of breast cancer and the final pathology and compare these results with those of other imaging modalities. This retrospective study included 925 women (mean age, 55.3 years ± 11.6) with 936 breast cancers, who underwent digital mammography, breast ultrasound, and magnetic resonance imaging before breast cancer surgery. AI-based tumor size measurement was performed on post-processed mammographic images, outlining areas with AI abnormality scores of 10, 50, and 90%. Absolute agreement between AI-based tumor sizes, image modalities, and histopathology was assessed using intraclass correlation coefficient (ICC) analysis. Concordant and discordant cases between AI measurements and histopathologic examinations were compared. Tumor size with an abnormality score of 50% showed the highest agreement with histopathologic examination (ICC = 0.54, 95% confidential interval [CI]: 0.49-0.59), showing comparable agreement with mammography (ICC = 0.54, 95% CI: 0.48-0.60, p = 0.40). For ductal carcinoma in situ and human epidermal growth factor receptor 2-positive cancers, AI revealed a higher agreement than that of mammography (ICC = 0.76, 95% CI: 0.67-0.84 and ICC = 0.73, 95% CI: 0.52-0.85). Overall, 52.0% (487/936) of cases were discordant, with these cases more commonly observed in younger patients with dense breasts, multifocal malignancies, lower abnormality scores, and different imaging characteristics. AI-based tumor size measurements with abnormality scores of 50% showed moderate agreement with histopathology but demonstrated size discordance in more than half of the cases. While comparable to mammography, its limitations emphasize the need for further refinement and research.