A new preoperative deep learning model can predict disease-free survival in patients set to undergo surgery for a common type of lung cancer.
That’s what South Korean researchers report after retrospectively studying more than 900 patients, publishing their findings May 12 in Radiology. Their artificial intelligence tool was trained to automatically extract prognostic information—invisible to the human eye—from CT scans without manual involvement.
With further testing and validation, Hyungjin Kim, with Seoul’s National University College of Medicine, and colleagues believe radiologists may one day use the tool to individualize treatment and achieve better outcomes.
As it stands, the recurrence rate for patients with early-stage, non-small cell lung cancers is “substantial,” ranging from 15%-38.5%, the group noted. This is despite technological advancements, such as endoscopy, that allow clinicians to better analyze lymph nodes prior to surgery.
Given these shortcomings, Kim et al. trained, tuned and internally validated their platform on 800 patients with all stages of adenocarcinoma over a six-year period.
Another separate dataset of 108 individuals was used for external validation, and carefully matched clinical factors such as site of disease, nodal disease, other tumors, types of surgery performed, and smoking history. It is a small, but robust dataset that will need to be expanded in future studies, said related editorialist Kitt Shaffer, with Boston University School of Medicine.
After clinical adjustments, the model independently predicted disease-free survival. Kim et al. noted that smoking status and deep learning model output were the only independent prognostic factors for such survival.
“The model output can be used as a preoperative risk stratification tool for surgical candidates with clinical stage 1 lung adenocarcinomas and may facilitate clinical decision making in the era of precision medicine,” Kim and colleagues concluded.
Shaffer pointed out that trusting AI may come slower for some radiologists than others, but said we should “welcome this method with open minds, but be vigilant about potential confounding factors.
“While it can be a bit unsettling to think that the computer is analyzing what we as radiologists cannot see or put intuitive names to, it would behoove us to become comfortable with this,” she added later. “We want to give our patients the full benefit of all of the information we obtain with our studies, particularly those involving ionizing radiation.”