An artificial intelligence (AI) tool developed at the University of Pennsylvania can help doctors predict the cancer risk in lung nodules seen on CT, according to a new study published in the journal Radiology. Pulmonary nodules appear as small spots on the lungs on chest imaging. They have become a much more common finding as CT has gained favor over X-rays for chest imaging.
“We see lung nodules in upwards of 20% to 30% of CT scans that we do,” said study senior author Anil Vachani, MD, MSCE, director of clinical research in the section of Interventional Pulmonology and Thoracic Oncology at the Perelman School of Medicine, University of Pennsylvania in Philadelphia. “So for anyone getting a CT scan for any reason, there is a one in three chance of discovering a pulmonary nodule.”
While the large majority of lung nodules are benign, a small minority are lung cancer. “With our current technology, it can become complicated if the nodule warrants more aggressive investigation like a biopsy, additional imaging testing like PET scans which deliver higher radiation doses, or to do surveillance,” he added. For this reason, interest in using AI or machine learning technologies has grown. “Tool like these, which look at the nodule on a CT scan and extract other features that we can’t do by the human eye, may be able to tell us more about whether something is more or less likely to be cancer,” he explains.
To evaluate the technology, Vachani and colleagues evaluated an AI-based computer-aided diagnosis tool developed by Optellum Ltd. of Oxford, England, to assist clinicians in assessing pulmonary nodules on chest CT.
In the study, six radiologists and six pulmonologists made estimates of malignancy risk for nodules using CT imaging data alone. They also made management recommendations such as CT surveillance or a diagnostic procedure for each case without and with the AI tool.
A total of 300 chest CTs of indeterminant pulmonary nodules were used in this retrospective study—150 with a known lung cancer, the other half were non-cancerous. Analysis showed that use of the AI tool improved the readers’ estimation of nodule malignancy risk on chest CT. Readers’ average AUC improved from 0.82 to 0.89. Each reader’s score improved but to varying degrees. The team also observed improved agreement among the different readers for both risk stratification and management recommendations.
Diving deeper into the data, Vachani found that for patients whose initial risk of cancer was below 5% based on existing guidelines in pulmonary medicine, analysis with the AI improved the sensitivity and specificity nearly to 100%. A similar trend in increased sensitivity and specificity was revealed in higher-risk patients. “It really helped identify those that are more likely to be cancer in this study,” Vachani said. “It could help expedite making a timely diagnosis and prioritize the most appropriate next steps for evaluation.”
Going forward, Vachani hopes to see this or another AI or machine learning tool studied in a prospective trial. “It will be important to see how a doctor uses it in a real clinical setting and observe how they manage patients going forward,” he recommended. “If that is accurate, it will help doctors and patients make better decisions.” In the meantime, the Optellum diagnosis tool is already FDA-approved and available for clinical use now.
