Victor Dahlblom
Doctoral student
Improving breast cancer screening with artificial intelligence
Author
Summary, in English
Introduction: The current standard method for breast cancer screening is digital mammography (DM). Digital breast tomosynthesis (DBT) can detect more cancers but is more resource-demanding, not the least due to a more time-consuming reading, which hinders the implementation in screening. Artificial intelligence (AI) might open possibilities to overcome this, but different potential ways of using AI need to be tested using representative screening data. To facilitate the testing and further development of AI, it is necessary to collect and organise more data in a research-friendly form.
Aim: To create a breast imaging research database and explore different ways of using AI to improve breast cancer screening.
Methods: All DM and DBT examinations performed in Malmö, Sweden during 2004–2020 were collected and combined with other relevant information in a research database. A subset consisting of 14 848 women had been examined with paired DM and DBT as part of the Malmö Breast Tomosynthesis Screening Trial (MBTST). This cohort was used to test different ways of using an AI cancer-detection system, which scores examinations based on cancer risk. It was studied whether the AI system could be used on DM to exclude normal cases from human reading, detect additional cancers on DM that radiologists only detected on DBT, or add DBT in selected high-gain cases. Further, it was studied how the AI system can be utilised to reduce the workload of DBT screening.
Results: A research database was created that contained 449 000 examinations from 103 000 women, performed during a time span of 17 years. This includes 9 250 cancers in 7 371 women. It was found that the tested AI system can be used on DM to exclude 19% of examinations from human reading without missing any cancers and that AI can detect 44% of DBT-only detected cancers using only DM. Further, adding DBT for the 10% of the women with the highest AI risk score can detect 25% more cancers than DM screening. For DBT screening, the AI system can reduce the reading workload to the level of DM screening, either by replacing the second reader in a double reader setup or by discarding half of examinations from reading, thus focusing double reading on the half with the highest risk.
Discussion: The results indicate that AI can be used to improve the performance and efficiency of breast cancer screening in several ways, including making it possible to use DBT in screening without demanding more resources. The research database can facilitate larger retrospective studies on these and other subjects. However, before clinical implementation, prospective studies would also be necessary, where e.g. the interaction between radiologists and AI can be investigated.
Aim: To create a breast imaging research database and explore different ways of using AI to improve breast cancer screening.
Methods: All DM and DBT examinations performed in Malmö, Sweden during 2004–2020 were collected and combined with other relevant information in a research database. A subset consisting of 14 848 women had been examined with paired DM and DBT as part of the Malmö Breast Tomosynthesis Screening Trial (MBTST). This cohort was used to test different ways of using an AI cancer-detection system, which scores examinations based on cancer risk. It was studied whether the AI system could be used on DM to exclude normal cases from human reading, detect additional cancers on DM that radiologists only detected on DBT, or add DBT in selected high-gain cases. Further, it was studied how the AI system can be utilised to reduce the workload of DBT screening.
Results: A research database was created that contained 449 000 examinations from 103 000 women, performed during a time span of 17 years. This includes 9 250 cancers in 7 371 women. It was found that the tested AI system can be used on DM to exclude 19% of examinations from human reading without missing any cancers and that AI can detect 44% of DBT-only detected cancers using only DM. Further, adding DBT for the 10% of the women with the highest AI risk score can detect 25% more cancers than DM screening. For DBT screening, the AI system can reduce the reading workload to the level of DM screening, either by replacing the second reader in a double reader setup or by discarding half of examinations from reading, thus focusing double reading on the half with the highest risk.
Discussion: The results indicate that AI can be used to improve the performance and efficiency of breast cancer screening in several ways, including making it possible to use DBT in screening without demanding more resources. The research database can facilitate larger retrospective studies on these and other subjects. However, before clinical implementation, prospective studies would also be necessary, where e.g. the interaction between radiologists and AI can be investigated.
Department/s
- Radiology Diagnostics, Malmö
- LUCC: Lund University Cancer Centre
Publishing year
2024
Language
English
Publication/Series
Lund University, Faculty of Medicine Doctoral Dissertation Series
Issue
2024:36
Full text
Document type
Dissertation
Publisher
Lund University, Faculty of Medicine
Topic
- Radiology, Nuclear Medicine and Medical Imaging
Keywords
- breast cancer
- artificial intelligence
- screening
- mammography
- breast tomosynthesis
- bröstcancer
- artificiell intelligens
- screening
- mammografi
- brösttomosyntes
Status
Published
Research group
- Radiology Diagnostics, Malmö
Supervisor
- Sophia Zackrisson
- Anders Tingberg
- Magnus Dustler
ISBN/ISSN/Other
- ISSN: 1652-8220
- ISBN: 978-91-8021-529-9
Defence date
5 April 2024
Defence time
09:00
Defence place
Rum 2005/2007, Carl-Bertil Laurells gata 9, vån 2, Skånes Universitetssjukhus i Malmö
Opponent
- Matthias Dietzel (Professor)