AMARA: Accurate MAlignancy Risk estimation of incidentally and screen-detected pulmonary nodules using Artificial intelligence

Problem description:

The Dutch-Belgian NELSON trial recently showed that CT screening leads to a 24% lung cancer mortality reduction in male long-term smokers. This has sparked discussions on how to implement lung cancer screening in the Netherlands. Screening using low-dose CT is highly sensitive for pulmonary nodule detection. The vast majority of these nodules are benign, but may result in extra investigations. Pulmonary nodules are also an increasingly common finding in routine medical care, with an incidence that is much greater than recognized previously. Thus, there is a critical need for improved nodule stratification and management.

Project summary

Deep learning, a methodology where computers can learn high dimensional features from large amounts of data, has led to a revolution in the field of artificial intelligence (AI) because its use resulted in major improvements in the performance of AI systems. Radboudumc has developed a deep learning-based algorithm for nodule malignancy prediction and results have shown that the algorithm performs comparable to thoracic radiologists for risk estimation of screen-detected nodules on low-dose chest CT.

To make the deep learning algorithm also applicable in routine clinical practice, the training data for the algorithm needs to be extended with CT data from routine clinical care. Therefore, an available large retrospective dataset of routine CT scan at Radboudumc will be curated and annotated. In addition, a more extensive external validation of the algorithm is needed. Therefore, we will perform an external validation of the algorithm on data from three European lung cancer screening trials: the Dutch-Belgian NELSON screening trial, the Danish Lung Cancer Screening Trial (DLCST), and the Multicentric Italian Lung Detection (MILD) trial, and on a large multi-center retrospective dataset collected at the five participating Dutch institutes.

We hypothesize that the final AI algorithm for nodule malignancy prediction and its integration in nodule management protocols will 1) accelerate the time to diagnosis of malignant nodules, and 2) limit unnecessary nodule work-up

Funding

This research project is funded by the Dutch Cancer Society. The project is led by Colin Jacobs and is a collaboration with investigators from Radboudumc, University Medical Center Groningen, Erasmus MC, UMC Utrecht and MUMC+.

People

Colin Jacobs

Colin Jacobs

Assistant Professor

 Firdaus Mohamed Hoesein

Firdaus Mohamed Hoesein

 Rozemarijn Vliegenthart

Rozemarijn Vliegenthart

 Hester Gietema

Hester Gietema

 Erik van der Heijden

Erik van der Heijden

 Robin Cornelissen

Robin Cornelissen