Same-time Monitoring of AI Performance Reduction and Thresholds in Breast Cancer Screening (SMART-SCREEN)
Developing methods to identify for which mammograms artificial intelligence-based breast cancer detection models are unreliable
State-of-the-art artificial intelligence (AI) models for breast cancer detection in mammograms can match the performance of a radiologist. However, these models always give an interpretation, even for cases in which the mammograms are unsuitable. In SMART-SCREEN, a new collaboration has been formed between two university hospitals (Radboudumc and NYU Grossman School of Medicine) and an AI company (ScreenPoint Medical B.V.) to develop methods to identify cases for which the model will be unreliable, ensuring the safer deployment of AI.
The Dutch National Screening Program has decreased breast cancer mortality significantly, but breast cancer risk continues to increase, and the increasing shortage of radiologists makes execution increasingly difficult. The implementation of AI into breast cancer screening could be a solution by reducing the workload of screening radiologists while improving their performance. However, before an AI model can be introduced, it is necessary to ensure its performance is as expected even when conditions change, due to errors in image acquisition or changes of the mammography machines.
The objective of SMART-SCREEN is to develop methods to identify when AI interpretations are unreliable. For this, we aim to create and test two methods that will detect mammographic examinations that are too different from those used for training and for which the performance of the AI model is decreased. The first method will require an additional AI model and will evaluate the entire breast, while the second method will be built-into the detection model but will only consider suspicious regions. These methods would increase the performance and trust in AI models and aid in the safe deployment of AI, even when circumstances change over time.
The deliverables of this project include multiple mammography datasets with varying image quality on which a trained AI breast cancer detection model was evaluated, and two evaluated methods to identify cases with reduced AI performance.
Andere projecten
In elk PPS-project werken kennisinstellingen, bedrijven en zorgpartijen samen aan een concrete gezondheidsvraag. Bekijk het volledige projectoverzicht.
Bekijk de programma’s
Binnen onze programma’s werken bedrijven, kennisinstellingen en regionale partners samen aan innovaties die bijdragen aan betere gezondheid en zorg.
Bekijk de partnerschappen
Strategische publiek-private partnerschappen (PPP's) zijn meerjarige samenwerkingen tussen bedrijven, kennisinstellingen en andere partijen.
Hoe kunnen we je helpen?
Heb je een vraag of wil je meer weten over Health Holland? We brengen je graag in contact met de juiste mensen, kennis en mogelijkheden om gezondheidsinnovatie te versnellen.
We hebben uw bericht ontvangen
We nemen binnen 3 tot 4 werkdagen contact met je op. Is je vraag dringend? Neem dan gerust direct contact met ons op. Hieronder zie je de contactoptie die het beste past bij je onderwerp.
Opnieuw beginnen
Er ging iets mis bij het versturen van het formulier. Probeer het opnieuw.