AIforScreening: Robust and trustworthy AI for breast cancer screening with mammography

AIforScreening: Robust og pålitelig KI for brystkreftscreening. Dette forskningsprosjektet undersøkte metoder og tilnærminger for å gjøre kunstig intelligens (KI) robust og pålitelig i tolkning av mammogrammer i brystkreftscreening. Prosjektet ble ledet av Norsk Regnesentral, med eksperter innen bildeanalyse og KI, og ble gjennomført i samarbeid med et team av medisinske eksperter fra Kreftregisteret, de norske brystsentrene og Karolinska Universitetssjukhuset. Prosjektet utviklet metoder for å trene KI-modeller til å gi stabile og pålitelige vurderinger av mammogrammer fra screening for brystkreft. Forklarbar KI var en sentral dimensjon: Modellene viste hvor i brystet de så tegn som kunne tyde på kreft, slik at radiologer lettere kunne vurdere grunnlaget for anbefalingene og fange opp situasjoner der modellen kunne ta feil. De samme mekanismene ble også brukt av modellene selv til å zoome inn på relevante områder i bildene. Ved å trene på millioner av norske mammogrammer oppnådde prosjektet resultater på nivå med de beste kommersielle produktene på markedet. Prosjektet studerte også betydningen av endringer i mammografibilder fra én undersøkelse til den neste, og utviklet metoder for å sikre at modellene fungerte godt på bilder fra andre mammografiapparater enn de var trent på.

Scientific impact: We developed machine-learning methodologies based on a transformer architecture that learned to zoom in on potential cancerous regions of a mammogram using only image-level cancer-diagnosis annotations. Our resulting cancer-classification models achieved AUC performance in the high 97% range for retrospectively classifying screen-detected cancers, which is high by industry standards. We implemented successful techniques for domain adaptation, improving performance on images from manufacturers not seen during training. We also studied the effect of change from one examination to the next, assuming that change would indicate a growing cancer. This did not significantly improve model performance, which may indicate a ceiling effect. Societal impact: We are currently in the process of evaluating the commercial potential of our model, and possible paths for utilizing it in the Norwegian mammography screening program. Regardless of the outcome of this process, the project has contributed to the awareness in the breast radiology community of the potential of AI to improve and support breast screening in addition to building national expertise in this important research field.

This research project will develop AI methods and approaches for robust, sustainable and trustworthy breast cancer screening with mammography. We expect findings will be transferable to other medical screening programs. Breast cancer is the most common cancer and the leading cause of cancer death among women worldwide. Early detection of breast cancer through screening is recommended by international health organizations to reduce this mortality. It is considered to expand the screening program by examining more women. The use of AI can lead to a larger capacity to do screening and detect cancers, reduce overdiagnosis, and by this save lives and give new opportunities for improved health services. Furthermore, we need to understand the effects of using AI in this context. To address this, the project will involve AI experts, medical experts, organizers and practitioners within mammographic screening to create a strong interdisciplinary team. The main novelties of the proposed project are: * Develop methods that can handle domain shifts without costly annotation of new data. Current methods have problems handing images from different equipment, operators and cohorts. * Exploitation of time series in AI breast cancer screening. Radiologists exploit prior mammograms when doing their manual reading, while it is rarely used in current AI-based models. * Interpretation of predictions from AI breast cancer models such that radiologists can understand the predictions. This gives trust and makes it easier to improve the AI method and combine this with manual reading by radiologists. * Find the differences in the prediction between radiologist and AI both statistically and using data collected in the screening program as a basis when we combine the two methods. * Develop AI systems that is usable for radiologists in mammographic screening utilizing the strength of each approach. A new screening program should give better predictions and use less radiology resources.