Artificial Intelligence Impact on Radiologist Performance in Breast Cancer Screening with Digital Breast Tomosynthesis.
Authors
Affiliations (4)
Affiliations (4)
- Baptist Health South Florida. Electronic address: [email protected].
- Manhattan Beach, California, USA.
- Baptist Health South Florida.
- JLS Consulting, Jupiter, Florida, USA.
Abstract
Assess impact of artificial intelligence (AI) on radiologists' detection of cancer on digital breast tomosynthesis (DBT) exams based on density, size, stage, and histopathology. Retrospective analysis of mammography audit data and screening cancers was conducted at four sites during two time periods with nine dedicated breast radiologists. Data was collected from March 1, 2018 to February 29, 2020 (pre-AI) and March 1, 2020 to February 28, 2022 with concurrent use of AI detection (post-AI). Age, density, tumor size, staging and histopathology were collected for all screen detected cancers. Endpoints were cancer detection rate (CDR), recall rate, tumor size, stage and histopathology. Pre-AI period had 54,440 exams (339 true positives) and post-AI had 48,742 exams (369 true positives). CDR/1000 improved from 6.23 to 7.57, increase of 1.34 (95%CI:0.33,2.36,p<.01). Recall rate was 6.97% pre-AI, post-AI was 6.96%, a decrease of 0.01% (95%CI:-0.32%,0.30%,p=0.47). Radiologists detected more cancers in dense breasts post-AI 45.0% vs. 37.2% pre-AI, increase of 7.8% (95%CI:0.6%,15.0%,p=.04). The mean size of invasive cancers in mm, pre-AI was 12.16, decreased to 10.74 post-AI, 1.42 mm smaller (95%CI:-2.83,-0.01,p<.05). More cancers were T-1 post-AI (70.7%) than pre-AI (63.1%), increase of 7.6% (95 CI:0.68%,14.53%,p=.03) without change in detection of DCIS post-AI (23.0%) vs. pre-AI (25.7%), p=.42. Invasive cancer detection rate/1000 increased from 4.63 pre-AI to 5.83 post-AI, difference 1.20 (95%CI:0.31,2.08,p<.01). Lobular cancers rate/1000 increased post-AI from 0.44 to 0.98, difference of 0.54 (95%CI:0.21,0.87,p<.001). Interpretation of screening DBT exams by dedicated breast radiologists with concurrent use of AI increased CDR by 22%, invasive detection rate by 26% and lobular detection rate doubled, an increase in cancers detected in dense breasts and a decrease in mean invasive size and stage. Use of AI was able to detect more invasive cancers without an increase in non-invasive cancers or recall rate.