Clinical value of lesion detection by using artificial intelligence system and reports for digital mammography
10.13929/j.1003-3289.201909006
- Author:
Tingting HA
1
Author Information
1. Department of Radiology, Peking University Shougang Hospital
- Publication Type:Journal Article
- Keywords:
Calcinosis;
Deep learning;
Mammography
- From:
Chinese Journal of Medical Imaging Technology
2019;35(12):1789-1793
- CountryChina
- Language:Chinese
-
Abstract:
Objective: To explore the clinical value of lesion detection artificial intelligence (AI) system based on deep learning (DL) for digital mammography. Methods: The mammograms and corresponding diagnostic reports of 484 patients were retrospectively analyzed. The sensitivity of AI system was evaluated in patients with Imaging-Reporting and Data System (BI-RADS) 3 and above lesions, χ2 and the consistency across different BI-RADS categories were observed. Among patients with BI-RADS 1 and 2 lesions but AI system indicating positive findings, further validation were performed by 3 attending radiologists, and the extra findings of AI were statistically analyzed according to BI-RADS scoring and types of lesions. Results: There were 103, 79, 23, 40 and 11 lesions categories with BI-RADS 3, 4a, 4b, 4c, 5, respectively, and the sensitivity of AI system was 82.52% (85/103), 97.47% (77/79), 100% (23/23), 100% (40/40) and 100%(11/11), respectively, with the overall sensitivity of 92.19% (236/256). No significant difference was found between report findings and AI findings across lesion categories (calcification, mass, asymmetry and distortion) nor BI-RADS categories (all P>0.05). AI system proposed 203 extra findings out of 254 patients. Validated by 3 attending radiologists, 75 patients with 80 lesions were categorized BI-RADS 0 (requires further information), and 21 patients with 23 lesions were categorized BI-RADS 3 and above. There was no statistically different among different type lesions with categorized BI-RADS 3 and above (all P>0.05). Conclusion: AI system has respectable sensitivity for finding lesions with BI-RADS 3 and above, therefore having the potential to reduce missed diagnosis during clinical practice.