NELSON-POP: Multi-source data approach for Personalized Outcome Prediction in lung cancer screening

Background

Lung cancer is the leading cause of cancer-related death worldwide, for which the five-year survival rates have yet to surpass 20%. Tobacco smoking remains the main risk factor for lung cancer. Although there is a decreasing prevalence of smokers in most countries, tobacco control is not the only measure for decreasing lung cancer mortality. In 2011, the National Lung Screening Trial (NLST) was the first multicenter randomized controlled trial (RCT) to demonstrate that three rounds of annual screening of a high-risk population using low-dose chest computed tomography (CT) lead to 20% fewer lung cancer deaths after seven years of follow-up, compared to annual screening with chest radiography. Over 53,000 participants were included in this landmark study. The Dutch-Belgian NELSON trial – the second largest RCT with 15,789 participants – recently published their results and showed a 24% mortality reduction in a high-risk population of men compared to no screening.

Problem description

If a population were invited for screening based on age and long-term smoking behaviour (like in NELSON), many individuals would not benefit. Many screenees have a low risk of lung cancer despite their smoking history (in NELSON <4% developed lung cancer), while some screenees have insufficient life expectancy to benefit from screening. Secondly, screen-detected lung nodules lead to extra investigations in 25-30% of screenees while most are benign. Thus, there is a critical need for stricter selection of screenees who will benefit from screening, and for improved nodule stratification.

Project summary

In this project, Radboudumc focuses on improving the efficiency of lung cancer screening by using artificial intelligence (AI). Our work is part of a larger consortium project: NELSON-POP. In this consortium, the unique expertise and data from the various NELSON sites and associated research groups are combined to leverage various unexplored data sources, in order to identify the factors most predictive of lung cancer. Using multi-source data, the consortium aims to maximize lung cancer screening efficiency, by developing prediction models to 1) optimize screenee selection, and 2) limit unnecessary nodule work-up.

We focus on the use of existing AI algorithms developed in our group for lung cancer screening. AI algorithms based on deep learning have great potential to perform more reproducible and more objective pattern recognition and thereby may increase the accuracy and consistency of malignancy probability estimation of pulmonary nodules. This increased accuracy can be used to develop optimized follow-up protocols, leading to less unnecessary follow-up CTs and unnecessary referrals in lung cancer screening. Therefore, the aim of this project is to accurately determine the probability of lung cancer of screen-detected pulmonary nodules using artificial intelligence in order to reduce the number of unnecessary repeat scans and unnecessary referrals, all contributing to reduction of radiation exposure, financial expenses, workload, invasive procedures, and screenee anxiety.

Funding

This consortium project is funded by the Dutch Cancer Society. The project is a public-private partnership and is led by prof. dr. Rozemarijn Vliegenthart (University Medical Center Groningen) and brings the expertise of University Medical Center Groningen, Erasmus MC, UMC Utrecht, KU Leuven, Radboud UMC and MUMC+ together. The consortium project is co-funded by Siemens Healthineers.

People

Ernst Scholten

Ernst Scholten

Radiologist

Mathias Prokop

Mathias Prokop

Professor

Radboudumc

Colin Jacobs

Colin Jacobs

Assistant Professor