Automated classification/ detection of Incidental Findings in LCS scans
During LDCT screening, potentially significant abnormalities unrelated to lung cancer, known as incidental findings (IFs), are often detected. These can include conditions such as coronary artery calcification, emphysema, and interstitial lung diseases. In LCS trials, IFs were reported in 33.8% to 45% of participants. The high prevalence of these IFs has led to guidelines ensuring their reporting and management are evidence-based and cost-effective. However, detecting IFs often leads to additional diagnostic interventions, increasing screening costs and risks.
AI models in LCS have been extensively studied for lung nodule detection and malignancy risk estimation. However, radiologists also need to detect significant IFs, and omitting these findings can harm patients and have medicolegal implications. Therefore, AI models must also detect IFs to advance towards autonomous LCS scan interpretation.
This project aims to automate the classification and detection of IFs from LCS CT scans. It involves comparing existing AI algorithms for detecting/classifying IFs, including narrow AI models for specific IFs and broader models for multiple IFs. Additionally, the project will develop and/or fine-tune AI algorithms for IF classification/detection in a lung cancer screening setting. We plan to extract IF annotations from existing annotations and manually annotate a subset of the data to create a benchmark dataset for algorithm evaluation. This will help establish a reliable dataset to assess and improve AI algorithms for IF detection in LCS. The developed AI algorithms will also be evaluated for fairness.
The main aim is the automated classification and/or detection of IFs in LCS CT scans.
a) To automatically label unstructured text data collected during the NLST, to identify key findings.
b) To develop and/or finetune AI algorithms for incidental finding classification/ detection in a lung cancer screening setting.
c) To compare existing AI algorithms for the detection/ classification of Incidental Findings in a Lung Cancer Screening setting. These algorithms can be narrow AI algorithms (designed for the purpose of detecting one specific Incidental Finding, i.e. Emphysema) or broader AI algorithms, such as different foundation models that can detect multiple incidental findings.
d) To evaluate the fairness of AI algorithms in lung cancer screening, to determine if there is varying performance on detection, classification tasks for different subpopulations (sexes, ages, ethnicities, socioeconomic backgrounds; dependent on the available data).
Lisa Klok, Co-Investigator, Department of Medical Imaging, Radboud University Medical Center Nijmegen, Netherlands
Fennie van der Graaf, Co-Investigator, Department of Medical Imaging, Radboud University Medical Center Nijmegen, Netherlands
Lena Philipp, Co-Investigator, Department of Medical Imaging, Radboud University Medical Center Nijmegen, Netherlands
Michel Vitale, Co-Investigator, Department of Medical Imaging, Radboud University Medical Center Nijmegen, Netherlands
Colin Jacobs, PI, Department of Medical Imaging, Radboud University Medical Center Nijmegen, Netherlands
Mathias Prokop, Department of Medical Imaging, Radboud University Medical Center Nijmegen, Netherlands