Operational insights into AI-assisted mammographic interpretation for enhanced sono-mammography performance

Abstract Background Breast cancer is the major cause of cancer-related mortalities among females worldwide. Digital mammography often coupled with ultrasound is the first line to detect and diagnose different breast diseases. Recently AI was introduced as a tool to enhance breast cancer screening. Objectives This study aimed to evaluate the added value of combining mammographic artificial intelligence to baseline sono-mammography in the primary diagnosis of different breast lesions. Methods In total, 108 patients with 133 breast lesions were included in our study. These patients had full field digital mammography (FFDM) and ultrasound (US) examination. The mammogram pictures were analyzed using AI software system (Lunit INSIGHT, Korea, for Fujifilm digital mammography system), where each lesion was given an abnormality score reflecting probability of malignancy. The ACR-BIRADS results of FFDM and US examinations as well as AI results were compared to the histopathological analysis (for suspicious lesions) or close follow-up (for benign looking lesions) as the gold reference standard. Results Sono-mammography had a sensitivity of 95.9%, a specificity of 92.9% with a diagnostic accuracy of 93.98%. Correlating the AI scoring values to the BIRADS categories and hence the probability of malignancy was statistically significant among all categories (P-value < 0.001). According to AI, a cutoff value of 36% was established to differentiate between benign and malignant cases. AI had a sensitivity of 87.8% and a specificity of 85.7% with a diagnostic accuracy of 86.46%. Upon combining the results of sono-mammography and AI, sensitivity was 100% and specificity was 81%. Combined diagnostic accuracy was found to be 88% which is superior to AI when used solely (85.7%), yet still inferior to sono-mammographic accuracy (93.9%). Conclusions The integration of AI-based interpretation with sono‑mammography improved sensitivity, but reduced specificity compared to the standard approach.