Factors Associated with False-Positive Marks and False-Negative Cancers on Automated Breast Ultrasound Using an Artificial Intelligence-Based Computer-Aided Detection System.

Artificial intelligence-based computer-aided detection (AI-CAD) systems have been developed to assist radiologists in interpreting automated breast ultrasound (ABUS), but factors associated with false-negative (FN) cancers and false-positive (FP) AI-CAD findings remain unclear. This single-center retrospective study evaluated stand-alone AI-CAD performance in ABUS and identified factors associated with FN cancers and FP findings. The study included 338 women with final negative screening ABUS results and 97 women with screening-detected breast cancer. AI-CAD sensitivity and specificity were evaluated. Characteristics of cancers with and without AI-CAD markers were compared, and patient-level logistic regression and lesion-level generalized estimating equations were used to identify factors associated with FP results and pseudolesions. AI-CAD detected 72 of 97 cancers, yielding a sensitivity of 74.2% and specificity of 57.7%. FN cancers were more often nonmass lesions (20.0% vs. 4.2%, P = 0.025), isoechoic masses (55.0% vs. 10.1%, P < 0.001), and masses without an echogenic rind (60.0% vs. 34.8%, P = 0.043) compared with AI-CAD-detected cancers. Age (odds ratio [OR], 0.94; 95% confidence interval [CI], 0.91-0.97; P < 0.001) and heterogeneous background echotexture (OR, 1.75; 95% CI, 1.09-2.82; P = 0.022) were associated with patient-level FP results. At the lesion level, dense breasts (OR, 2.02; 95% CI, 1.14-3.59; P = 0.017), heterogeneous background echotexture (OR, 2.30; 95% CI, 1.50-3.53; P < 0.001), posterior location (OR, 2.07; 95% CI, 1.15-3.70; P = 0.015), detection on the opposite side of the scanned view (OR, 2.47; 95% CI, 1.51-4.03; P <0.001), and number of detected views (OR, 0.06; 95% CI, 0.03-0.12; P < 0.001) were associated with pseudolesions. Specific lesion and imaging features are associated with FN cancers and FP AI-CAD findings on ABUS, suggesting that AI-CAD should be used as an adjunctive tool rather than as an independent screening reader.