This article is more than 5 years old. Information may no longer be current.
Variables may help differentiate malignant, benign lung nodules
Click Here to Manage Email Alerts
A novel radiomics-based approach to lung cancer screening with low-dose CT may help differentiate between benign and malignant lung nodules, according to study results.
Tobias Peikert, MD, a pulmonologist and critical care specialist at Mayo Clinic in Rochester, Minn., and colleagues on a multi-institutional, multidisciplinary research team used features such as sphericity, flatness, elongation, spiculation, lobulation and curvature to determine whether 726 indeterminate nodules (all 7 mm)were malignant or benign.
The data set included 318 benign nodules and 408 malignant nodules.
A multivariate model selected eight of 57 quantitative radiologic features to differentiate between the two. The features included vertical location (offset carina centroid z); volume estimate (minimum enclosing brick); flatness and density, with a texture analysis (Score Indicative of Lesion/Lung Aggression/Abnormality); surface complexity (maximum shape index and average shape index); and estimates of surface curvature (average positive mean curvature and minimum mean curvature).
All features carried a significant association (P < .01 for all), along with an area under the curve of 0.939.
HemOnc Today spoke Peikert about the challenges differentiating between malignant and benign nodules, the value of the criteria he and his co-investigators identified, and whether this approach may become standard.
Question: Why is it so difficult in lung cancer screening to differentiate between malignant and benign nodules?
Answer: The trigger for this research was the publication of the National Lung Screening Trial (NLST) in 2011. The good news about the screening trial was that it showed low-dose CT on an annual basis actually decreased all-cause mortality and lung cancer-specific mortality in the context of detecting more tumors and treating them at an earlier stage. However, the remaining challenges with lung cancer screening are still numerous. One of the biggest is that in the NLST study, 96% of nodules 4 mm in diameter or larger in higher-risk individuals actually turned out to be benign. That represents a 96% false positive rate. When patients are told they have a lung nodule, it causes a lot of anxiety and distress. As clinicians, we will follow most of these nodules using repeated CT scans. Although CT is not invasive, it is problematic in terms of radiation exposure and contributes to health care costs. In addition, a subgroup of these patients may also be exposed to tissue biopsies, additional imaging such as PET scans, or surgical procedures, which also expose the patient to more undue iatrogenic risk — including complications or even mortality — and unnecessary health care costs. With all of this in mind, we had to come up with a better way to determine whether these nodules are benign or malignant. When we see patients with lung nodules in the clinic, we use one of several clinical risk prediction models. These online calculators use clinical data — such as age, smoking history, cancer history, number, size and location of the lung nodules, and other variables — determine the likelihood that a given nodule represents cancer. These calculators are mostly helpful to identify low-risk nodules, for which active surveillance with repeat CT typically is recommended, and high-risk nodules, for which surgical resection typically is recommended. Unfortunately this process varies significantly based on which model is used. Better tools are clearly needed, especially for intermediate-risk lung nodules. These are primarily nodules 7 mm or larger. That’s why we looked at the nodules with this size in the NLST study dataset.
Q: How did you conduct your study?
A: We included 408 malignant lung nodules and 318 benign nodules 7 mm or larger detected during the NLST study to develop a radiomics model to differentiate cancers from benign lesions. We started with 57 variables and narrowed them down to the eight features included in our current model. Using this model, we had excellent diagnostic accuracy and we were able to validate our results internally using the LASSO technique. The variables used are quite different from features used in clinical risk calculators and can be extracted automatically without input from the physician. The next step is to validate our model in another independent data set, and we are working on that. Although the preliminary data from the validation looks promising, we will have to wait for the final results. Hopefully, it will translate into a clinically useful test that could be applied to this problem.
Q: Is there no clinical decision-making involved ?
A: The model is currently a radiomic-based model. It’s in the research realm, but the way this works is that a radiologist would have to click on the nodule to identify it. Then the computer program segments the nodule and some of the surrounding tissue. The software does the analysis. It looks at these eight variables that are based on the location, shape, surface characteristics and the texture characteristics of the nodule. After the computer extracts this information, you get a decision of whether the nodule is more likely benign or malignant. The clinician would use this information in their discussion with the patient and integrate it into the clinical decision-making process. We use this information to clinically rule in or rule out cancer. If I am very confident this is a malignant nodule, I will send the patient for treatment, such as surgery or radiation, avoiding additional diagnostic testing, biopsy or PET/CT scans. The more challenging part is to use this as a rule-out test. In this case, you look at the nodule and determine that there is a nearly 100% chance that this is not a cancer, but a benign nodule, where you can rule out any further action. However, the rule-out strategy may also provide more confidence to following these patients with serial CT scans. Right now, our test is probably one of the best imaging biomarkers for this indication reported in the literature. The caveat is that we need to validate it externally. That’s where the rubber hits the road.
Q: So this validation cohort will look at intermediate-risk nodules, not the ones for which status can be confidently determined?
A: That’s right. That’s why we eliminated the nodules in the range of 4 mm to 7 mm. Current guidelines already define the minimum diameter of clinically significant pulmonary nodules as 6 mm. Again, though, this comes with the caveat that every cancer starts little, so you have to be very careful when you immediately throw a nodule out of consideration for cancer.
Q: Will the external validation use similar protocols as yours?
A: Yes. The benefit of using the screening NLST cohort as our training set is that these CT scans were obtained at large number of centers across the United States. This included scanners from different manufactures using slightly different protocols, which means this represents real-life data. Our validation cohort is from the DECAMP study, which also includes incidentally discovered pulmonary nodules. We are awaiting the final results of the study before we complete our blinded validation. If our validation is successful, we hope to apply our approach to both screen- and incidentally detected pulmonary nodules. – by Rob Volansky
References:
National Lung Screening Trial Research Team. N Engl J Med. 2011;doi:10.1056/NEJMoa1102873.
Peikert T, et al. PloS One. 2018;doi:10.1371/journal.pone.0196910.
For more information:
Tobias Peikert, MD, can be reached at Mayo Clinic, 200 1st St. SW, Rochester, MN 55905; email: peikert.tobias@mayo.edu.
Disclosures: The study was funded by the Department of Defense. Peikert reports no relevant financial disclosures.