Clinical application of novel mobile AI solution for real-time detection and differential diagnosis in breast ultrasound: The first prospective feasibility study.

person Jeeyeon Lee Hogukro 807, Buk-gu, Daegu, South Korea info_outline Jeeyeon Lee, Fiona Cheng Tsui-Fen

Authors person Jeeyeon Lee Hogukro 807, Buk-gu, Daegu, South Korea info_outline Jeeyeon Lee, Fiona Cheng Tsui-Fen Organizations Hogukro 807, Buk-gu, Daegu, South Korea, Shin Kong Wu Ho-Su Memorial Hospital, Taiwan, Taiwan Abstract Disclosures Research Funding No funding sources reported Background: Existing artificial intelligence (AI) solutions for breast ultrasound faced limitations due to their non-real-time detection and server dependency. However, novel mobile real-time AI solution now enable on-device detection and differential diagnosis, aiding in immediate decision-making. This study aims to evaluate the feasibility and effectiveness of such a mobile AI solution in clinical settings, by comparing it against expert breast imaging diagnoses. Methods: From August to December 2023, a feasibility study was conducted in tertiary medical center of Taiwan using mobile AI solution (CadAI-B for breast, BeamWorks Inc., Korea). CadAI-B is connected via HDMI or DVI to an ultrasound equipment using tablet PC. It supports healthcare providers to detect suspicious breast lesions by highlighting suspicious area during breast ultrasound scanning, and to make a differential diagnosis by providing BI-RADS categories and malignancy score (0-100%). The rate of real-time detection in malignancies and area under the receiver operating characteristic curve (AUC) was calculated. Distribution of BI-RADS categorized by CadAI-B were compared with those by breast imaging experts on independent review. Results: The analysis included 33 patients with an average age of 47 years (IQR, 44 – 62). There were 14 malignancies, 17 benign lesions, and 2 normal cases, and 30 patients (90.9%) underwent biopsy. For patients with malignancies, CadAI-B successfully identified all malignancies. Overall diagnostic performance, as AUCs calculated by the malignancy score and BI-RADS, was 0.835 and 0.850, respectively. The per-patient sensitivity of CadAI-B was 100.0%, while the specificity was 52.6%. The BI-RADS distribution between CadAI-B and expert was the same in malignant cases. In benign cases, CadAI-B categorized 9 cases (50.0%) as C4A or C4B, while experts classified 13 cases (72.2%), indicating a potential reduction in the need for biopsy. Conclusions: This study demonstrates that real-time processing capabilities of CadAI-B enable the identification of critical lesions during breast ultrasound, helping to reduce unnecessary biopsy. These findings may be helpful in diagnostic decisions and training of non-experts in clinical settings and offer the potential to improve workflow efficiency. Diagnosis BI-RADS categories by CadAI-B BI-RADS categories by expert 1/2 3 4A 4B 4C 5 1/2 3 4A 4B 4C 5 Normal (n = 2) 2 (100.0) 0 0 0 0 0 2 (100.0) 0 0 0 0 0 Benign (n = 18) 3 (17.6) 5 (29.4) 2 (11.8) 7 (41.2) 0 0 3 (17.6) 1 (5.9) 8 (47.1) 5 (29.4) 0 0 Malignancy (n = 14) 0 0 2 (14.3) 8 (57.1) 4 (28.6) 0 0 0 2 (14.3) 8 (57.1) 4 (28.6) 0 Note: The numbers in parentheses are the percentages.