Work

Initial results of an AI-guided evaluation of CE breast MRI

Year: 2025

Type: article

Abstract: The reliability of an AI-guided evaluation of breast MRI is tested on findings in the categories MR BIRADS 1 and 2 as well as MR BIRADS 4 and 5 categorized by well-experienced readers. An AI-pipeline was trained on images from nine clinical sites from seven countries. The pipeline automatically detects lesions in CE-MRI images, providing for each lesion a segmentation mask together with a lesion specific malignancy score on a scale from 0 to 100. The score was assigned to BIRADS categories for analysis. The CE breast MRI was evaluated in 78 patients (156 breasts) who were primarily assessed by experienced radiologists. The AI-guided evaluation consecutively included 78 single breasts in MR-BIRADS 1 or 2 category and 78 single breasts in MR-BIRADS 4 or 5 category, that were histologically clarified and revealed a benign or invasive tumour in stage pT1a,b or c. DCIS and invasive carcinomas over 20 mm were preliminary excluded. AI showed a true negative result in 72/78 breasts (specificity 92.3 %) in patients with normal or definitely benign findings (BIRADS 1,2), and false positive in 7.7 %. 44 invasive carcinomas (42 breasts) of stages pT1a (n = 10), pT1b (n = 12) or pT1c (n = 22) were correctly detected in 40 cases by the AI (sensitivity 90.9 %). The rate of false-negative findings in AI was 9.1 % (tumour sizes: 5 mm, 5 mm, 6 mm, 8 mm). AI would have avoided an intervention in 17 of 36 cases (47.2 %) with findings in MR-BIRADS 4 and 5 category, in which percutaneous biopsy revealed benign findings. The initial results of AI evaluation of breast MRI are promising. AI can become an important tool in the evaluation of breast MRI. • AI in MRI achieves a specificity of 92 % in normal breast or benign lesions. • AI in MRI achieves a sensitivity of 91 % in small invasive breast carcinomas. • AI allows a 46 % reduction in unnecessary biopsy procedures. • AI could become a tool for second opinion breast MRI.

Source: European Journal of Radiology Artificial Intelligence

DOI: 10.1016/j.ejrai.2025.100057