2.1. Dataset

This retrospective study analyzed data originally taken from 200 low-dose CT scans of the lungs of patients at the University of Iowa Hospital. Pathology and radiology reports were reviewed to identify an analysis set of patients who met eligibility criteria of having (a) a solitary lung nodule (5–30 mm) and (b) a malignant nodule confirmed on histopathology or a benign nodule confirmed on histopathology or by size stability for at least 24 months. Manual segmentations were performed by a graduate student trained in medical image analysis in order to define a region of interest (ROI) around each nodule. The ROIs were defined to include amounts of parenchyma approximately proportional to the nodule sizes. Individual ROI voxels were labeled as belonging to either the nodule or the parenchyma, with radiomic features calculated separately for each to produce the complete set of 416 (approximately half nodule and half parenchyma) quantitative imaging biomarkers. These biomarkers measured features such as intensity, shape, and texture of the ROI (15). This study is a secondary analysis of de-identified data originally collected with approval from the University of Iowa institutional review board. Demographic information can be found in Table 1. The dataset is available as Supplementary Material.

Demographics of patient cohort.

A strength of the dataset is its fairly balanced malignant/ benign status breakdown, with 45% of the cases malignant and 55% benign. Many machine learning-based classifying algorithms assume that the outcomes of a data set are balanced, but this assumption is not met when the proportion of outcomes is highly uneven. The data set used in this work has a nearly even ratio of malignant and benign nodules (16).