With an estimated 229,000 new cases and 136,000 deaths in the United States alone, lung cancer remains the deadliest malignancy worldwide (1). Recently, however, the NLST (National Lung Screening Trial) and the NELSON (Dutch-Belgian Randomized Lung Cancer Screening Trial) studies have demonstrated improved lung cancer mortality for low-dose computed tomographic (CT) screening of the chest in high-risk individuals, and, consequently, lung cancer screening programs are being implemented globally (2, 3). Although this is very exciting, numerous challenges remain, including the detection of large numbers of benign pulmonary nodules, diagnosis of indolent lung cancers, and many others. The implementation of lung cancer screening and the increased use of diagnostic chest CT, together with advances in CT technology, will undoubtedly lead to an ever-increasing number of detected lung nodules. An estimated 20 million chest CT scans are being performed annually in the United States alone (4, 5). Despite the reliance on predictive models and nodulemanagement practice guidelines, considerable variability in nodule classification and uncertainty in management remain (6, 7). Continued research exploring new biological and imaging-based biomarkers is crucial to meeting these challenges. In this issue of the Journal, Massion and colleagues (pp. 241–249) report the development and external validation of a novel, computer-aided, deep learning–based radiomic model, the Lung Cancer Prediction Convolutional Neural Network (LCP-CNN), to distinguish benign nodules from malignant screen-detected and incidentally detected indeterminate pulmonary nodules (8).

Radiomics refers to the identification, extraction, quantification, and analysis of imaging features from radiologic images, with the goal of better or more consistently characterizing radiologic findings. For lung nodules, quantitative and qualitative density and morphologic features provide objective characterization not available by standard visual image interpretation. The analysis of already-available imaging data renders this approach to development and validation of nodule radiomics safe and cost effective. In contrast to conventional radiomic methods in which imaging features are selected by experienced clinicians, deep learning–based radiomics relies on machine learning–extracted features that are frequently abstract and commonly difficult to link back to the underlying biology. Several other recent studies have explored the potential role of radiomics in the classification of indeterminate pulmonary nodules with promising results (9–11). Enthusiasm has, however, been tempered by the lack of consistency in radiomics features included in these models, the need for homogeneous image acquisition, a lack of stability of the imaging features, the small numbers of scans in relationship to the extracted imaging features (type I error), and a lack of external validation. Models derived from large, heterogeneous real-life data sets, such as the NLST, that are further validated in external data sets, as in the current study, are needed.