Impact on clinical outcomes of implementing artificial intelligence in digital mammography screening with double reading

Congress:

ECR 2023

Poster Number:

C-18246

Type:

Educational Exhibit

Keywords:

Artificial Intelligence, Breast, CAD, Mammography, Neural networks, Computer Applications-Detection, diagnosis, Education, Cancer, Education and training

Authors:

A. Gräwingholt, M. Peratikos, J. Hoffmeister, R. Gräwingholt, J. Shisler, A. Y. Toledano

DOI:

10.26044/ecr2023/C-18246

DOI-Link:

https://dx.doi.org/10.26044/ecr2023/C-18246

Learning objectives

To demonstrate the influence of artifical intelligence (AI) on double reading in mammographic screening with analysis of outcomes and example cases showing AI's impact on cancer detection.

Background

Five radiologists’ performance reading 2D digital mammograms (DM) at two sites was retrospectively evaluated during two time periods. Outcomes were assessed from July 1, 2009 to June 30, 2019 (pre-AI) and July 1, 2019 to August 31, 2020 (post-AI). The five radiologists all have more than 15 years’ experience in reading mammograms. AI (iCAD ProFound AI V2.0, high operating point) was used by two of the five radiologists, and during the mandatory consensus conference. The DM systems included 2 GE Essentials and 1 Siemens Inspiration....

Findings and procedure details

Performance rates were compared for women screened pre-AI (n=322,793; 1451 cancers) and post-AI (n=30,824; 174 cancers). CDR improved from 4.54 (pre-AI) to 5.78 (post-AI), a difference of 1.24 (95% CI: -0.03, 2.79). AIR changed from 4.56% (pre-AI) to 4.84% (post-AI), a difference of 0.28 (95% CI: -0.14 ,0.70). [Table 1] Individual CDRs all increased. [Fig 2] Example cases illustrate cancers detected due to use of AI.1 Case 1 [Fig 3] An asymptomatic 60-year-old woman presented for a subsequent round of screening. Reader with AI found...

Conclusion

Implementation of AI with double reading in DM based breast cancer screening increased CDR for the screening unit (two sites) on average by 27%. The increase of the average is mainly driven by the high CDR of those radiologists using AI as a supportive tool. The AIR was not appreciably increased. These results are consistent with prior studies and the conditional recommendation in the EU guidelines2 to use AI as a supportive tool in double reading for mammographic screening.

Personal information and conflict of interest

A. Gräwingholt: Advisory Board: iCAD, Inc. M. Peratikos: Consultant: iCAD, Inc. J. W. Hoffmeister: Employee: iCAD, Inc. Shareholder: iCAD, Inc. R. Gräwingholt: Nothing to disclose J. Shisler: Consultant: iCAD, Inc. A. Y. Toledano: Consultant: iCAD, Inc.

References

Mammographiescreening Paderborn, case images, figures 3, 4, 5 https://healthcare-quality.jrc.ec.europa.eu/ecibc/european-breast-cancer-guidelines/artificial-intelligence/double-reading-AI-support

Fig 1: Reading Process for Screening Mammograms Flowchart

Table 1: Radiologist Performance Rates Pre and Post AI Implementation

Fig 2: Individual and Average Radiologist Performance Rates Pre and Post AI...

Fig 3: Case 1

Fig 4: Case 2

Fig 5: Case 3