09 Radiologists’ and Ultrasound Artificial Intelligence Decision-Support Assessment of Benign and Malignant Cystic Breast Lesions

Background

Cystic breast lesions, ranging from benign to malignant, are commonly encountered in clinical practice and may present diagnostic challenges. Accurate characterization is essential to guide management. We evaluated the diagnostic performance of an ultrasound artificial intelligence (AI) decision-support system compared with breast radiologists for classifying cystic breast lesions.

Materials and Methods

This retrospective study included 77 breast ultrasound cases from 2015 to 2025 at a single academic institution—50 consecutive biopsy-proven cystic carcinomas and 27 consecutive biopsy-proven benign cystic lesions. Patients were aged 18 years or older. The cystic carcinomas represented histologic subtypes that radiologists typically recognize as cystic breast malignancies, including mucinous, medullary, papillary, and metaplastic carcinomas. Purely solid lesions without cystic components were excluded. Demographics, clinical presentation, and imaging features were collected. All examinations were performed by breast-specialized sonographers using standardized protocols. Each lesion was interpreted by a fellowship-trained breast radiologist and retrospectively analyzed using the KOIOS DS™ AI decision-support tool. KOIOS classifications were benign, probably benign, suspicious, or probably malignant. Pathology served as the reference standard. Statistical analysis included 2-tailed tests and ANOVA, with diagnostic accuracy calculated; P <.05 was considered significant.

Results

Among 77 cystic breast lesions (50 malignant, 27 benign; Table), patients with cystic carcinomas were significantly older than those with benign lesions (66.73 vs 53.15 years, P <.05). Palpable masses were more frequent in benign cystic lesions (70.4%) than in cystic carcinomas (46%; P <.05). Malignant lesions more often exhibited irregular shape, non-circumscribed margins, non-parallel orientation, and vascularity (P <.05). Radiologists assigned BI-RADS 4/5 in 92% of cystic carcinomas and BI-RADS 2/3 in 56% of benign lesions. KOIOS correctly classified 76% of cystic carcinomas as suspicious/probably malignant and 26% of benign cystic lesions as benign/probably benign. Overall diagnostic accuracy was higher for radiologists compared with KOIOS. The difference was statistically significant for malignant lesions (P <.05), while the difference for benign lesions did not reach significance (P = .09).

Conclusion

Radiologists demonstrated superior diagnostic accuracy compared with the KOIOS DS™ AI tool in classifying malignant vs benign cystic breast lesions. A benign classification by the AI tool should not override the radiologist’s clinical judgment when imaging features indicate the need for biopsy.