By David Yankelevitz, MD
Posted: August 14, 2019
IN REFERENCE TO: Peikert T, Rajagopalan S, Karwoski RA, et al. Novel high-resolution computed tomography-based radiomic classifier for screen-identified pulmonary nodules in the National Lung Screening Trial. PLoS One. 2018;13(10). e0196910
The article “Novel high-resolution computed tomography-based radiomic classifier for screen-identified pulmonary nodules in the National Lung Screening Trial”1 outlines the tremendous need for developing techniques to distinguish benign from malignant nodules, especially small lesions. This is especially important with the endorsement for lung cancer screening now in place in multiple countries based in part on the positive NELSON trial; these results will surely lead to further uptake of screening.2 In addition, with the inclusion of the large number of incidental nodules found outside of the screening context, the authors describe a “potential emerging global epidemic of newly detected lung nodules.” With continued improvements in scanner technology and availability, as well as computer-assisted means for detecting small nodules, this challenge will surely continue to be in the forefront.
Tackling False Positives
One of the widely accepted challenges in screening (and also for the incidental nodule) has been what is described as the high rate of false-positive results. Nodules may require additional work-up, possibly leading to invasive procedures and their potential for harm; in some cases, these nodules turn out to be benign, in which case patients go through potentially unnecessary thoracic biopsies or explorations. Various nodule-management protocols have been developed for the purpose of minimizing these false positives, primarily using a combination of size thresholds for initiating work-up and then using growth estimates based on follow-up scanning. There is clearly a need to continue to make these evaluations more efficient.
The method outlined in the paper relies on the use of radiomics, a method of extracting features from images and determining their predictive value. For their analysis, the authors chose a dataset of nodules from the publicly available database of the National Lung Screening Trial, which was the first and largest of the trials to demonstrate a mortality reduction for lung cancer screening. The availability of this type of large, well-documented database is an important resource, as it will continue to facilitate these types of analyses well into the future.
The approach taken for their evaluation involved the analysis of 57 different features. These particular features were chosen specifically with a view toward incorporating ones already considered to have clinical significance. Using a variety of well-known statistical techniques, the authors were able to optimize their prediction model using only eight features, demonstrating an area under the curve of (AUC) of 0.94. This represents a highly promising result, although the authors suggest that additional validation using other datasets will be necessary.
However, when looking more closely at these results, a challenging aspect appears in that nearly all of the diagnostic information can be explained by nodule size alone. The AUC just using volume or other measures that reflect size was at least 0.9. Although the authors attempted to account for this by eliminating size-dependent measures and still show a high AUC, it seems likely that at least some of the remaining metrics remain size dependent.
Because size measurements provide so much of the diagnostic information, it is difficult to imagine that once size is accounted for when making a diagnostic consideration about a particular nodule that the additional small bit of information provided by other features would substantially change management. This point has been emphasized previously by Reeves et al.3 Several other considerations also dampen the enthusiasm for this approach. The first is the use of simple dichotomization when comparing benign versus malignant, as this does not account for the extensive variation within each of these categories. It seems likely that the different types of benign nodules (infectious, chronic infectious, and benign tumors) would have very different features; similarly, different types of malignant tumors with known differences in growth patterns would also prove quite different. Beyond that, the database analyzed was not representative of the distribution of nodule types in the screening population, with nearly 50% of the nodules chosen proving malignant—a point that the authors recognize. Additionally, the scan parameters in the National Lung Screening Trial database are already outdated, with slice thickness of 2.5 mm compared to modern protocols, which routinely obtain submillimeter slice thickness. Finally, one of the most important clinical pieces of information that greatly affects decision making about the type of nodule is whether it is identified in a baseline round or subsequent round of imaging; this was not explicitly accounted for in the analysis. Here the effect of size is in the opposite direction; for a baseline nodule, the larger the size the more likely it proves malignant, whereas for new nodules, the relationship is more complex, but after a certain threshold, increasing size implies a decreased likelihood of malignancy.
In conclusion, the overarching goal of this approach to identify additional features of “screened” nodules so as to better classify them is of great importance, and it seems likely that additional information can be captured using various radiomic features. However, it also seems likely that a more nuanced approach that incudes higher-quality images, a much larger database of cases with consideration for various types of nodules, improved feature selection, and further inclusion of additional clinical information will be needed before we can realistically change current approaches to nodule management. Nevertheless, all of these considerations can be addressed. This paper demonstrates the great potential for this strategy, which should continue to improve over time. ✦
About the Author: Dr. Yankelevitz is a professor of radiology and the director of the Lung Biopsy Service at the Icahn School of Medicine at Mount Sinai. Dr. Yankelevitz is a named inventor on a number of patents and patent applications related to the evaluation of diseases of the chest including measurement of nodules. Dr. Yankelevitz has received financial compensation for licensing of these patents. In addition, he is a consultant and co-owner of Accumetra, a private company developing tools to improve the quality of CT imaging and is on the medical advisory board for Grail a company that does blood tests for early detection of cancer.
1. Peikert T, Rajagopalan S, Karwoski RA, et al. Novel high-resolution computed tomography-based radiomic classifier for screen-identified pulmonary nodules in the National Lung Screening Trial. PLoS One. 2018;13(10). e0196910.
2. De Koning H, Der Aalst C, Ten Haaf K, et al. Effects of Volume CT Lung Cancer Screening Mortality Results of the NELSON Randomized-Controlled Population Based Trial. Presented at: IASLC 19th World Conference on Lung Cancer (Abstract PL02); September 25, 2018; Toronto, Canada.
3. Reeves AP, Xie Y, Jirapatnakul, A. Automated pulmonary nodule CT image characterization in lung cancer screening. Int J Comput Assist Radiol Surg. 2016;11(1):73-88.