Performance of Artificial Intelligence in Breast Ultrasound for Non-Mass Lesions

Background: In recent years, FDA-approved artificial intelligence (AI) decision-support systems have become commercially available, demonstrating excellent performance in the analysis of breast lesions. More recently, The ACR BI-RADS v2025 ultrasound lexicon newly defines a non-mass lesion (NML) as a discrete three-dimensional finding that differs from normal tissue but lacks the discrete margins and assignable shape of a mass. To date, the performance of AI in the evaluation of NML remains unclear.

Methods: In this retrospective single-center study, we evaluated the performance of an AI decision-support system in NML compared to masses. Hebrew breast-ultrasound reports were searched for terminology suggestive of NML, including equivalents of “undefined structure” and “undefined area.” Lesions with biopsy results or at least two years of imaging stability were included. Lesion images were processed using AI software (Koios DS), which generated a BI-RADS- score. Diagnostic performance was compared for the two groups, assigned in a 1:2 ratio, using two-sided Fisher exact tests.

Results: Fifty reported NMLs were identified and 100 reported masses were randomly collected for analysis. Koios achieved 100% sensitivity, 23.8% specificity, and 36.0% accuracy, compared with 85.7%, 81.4%, and 81.6% for masses, respectively. Specificity and accuracy were significantly lower for non-mass lesions (both p<0.001), whereas sensitivity was not significantly different (p=0.47). Binary management agreement between Koios and the radiologist was 80.0% for non-mass lesions and 72.0% for masses. After excluding non-mass cases in which relevant prior imaging or clinical history influenced the radiologist’s assessment, agreement increased to 92.9% (26/28).

Conclusions: AI maintained high sensitivity and provided management recommendations similar to radiologists for reported NMLs, particularly when cases influenced by prior information unavailable to the AI system were excluded. Nevertheless, its lower specificity highlights the importance of continuously reassessing AI performance as imaging terminology and clinical guidelines evolve.