Towards Fully Automated Volumetric Analysis of Lung Nodules in Computed Tomography

Abstract

Early detection of lung cancer significantly improves patient outcomes, and tracking the growth of lung nodules over time is key to understanding their progression and informing future treatment decisions. However, calculating nodule growth in computed tomography (CT) scans remains a highly manual and time-consuming task. In this work, we develop an automated end-to-end pipeline to compute lung nodule growth using state-of-the-art computer vision techniques. While modern advances in deep learning have all but solved many learning tasks in the domain of natural images, biomedical imaging presents unique challenges due to limited data availability, inconsistent annotations, and deployment constraints. We address these challenges by training robust detection and segmentation models using the LUNA16 and LNDb datasets. On the held-out UniToChest dataset, our methods generalize well, attaining a nodule recall of 77.49%, reducing false positives per scan by a factor of 11.3 compared to existing techniques, and achieving a mean nodule-wise Dice score of 0.6453. We then apply our methods to analyze nodule growth in 1,378 patients from the National Lung Screening Trial; we estimate a median nodule volume-doubling time of 791.23 days across all nodules from the patients that do not receive a cancer diagnosis and a median nodule volume-doubling time of 637.38 days across all nodules from the patients that do receive a cancer diagnosis. We also recall 82.20% of radiologist-annotated nodules that are directly associated with a cancer diagnosis and estimate a shorter median nodule volume-doubling time of 370.11 days for these nodules. By automating lung nodule growth quantification, this work lays the foundation for improved screening protocols, personalized treatment planning, and the development of novel imaging biomarkers. To encourage further work in this area, we release our full software pipeline at https://github.com/evanrubel/nodule_volumes.