AI can predict malignancy for multiple pulmonary nodules on Chest CT scans

AI can predict malignancy for multiple pulmonary nodules

By Erik L. Ridley, AuntMinnie staff writer

February 25, 2021 — An artificial intelligence (AI) algorithm was able to predict the risk of cancer in patients found to have multiple pulmonary nodules during CT lung cancer screening, according to research published online February 24 in Clinical Cancer Research.

A team of researchers from China and South Korea developed a machine-learning model that utilizes nodule characteristics on CT and pairs it with sociodemographic information to predict malignancy risk in patients with multiple pulmonary nodules. In testing, the algorithm — called PKU-M — yielded better results than previous models and also outperformed three thoracic surgeons and a radiologist in a prospective evaluation.

“Our prediction model, which was exclusively established for patients with multiple nodules, can help not only mitigate unnecessary surgery but also facilitate the diagnosis and treatment of lung cancer,” said co-author Dr. Young Tae Kim, PhD, of Seoul National University Hospital in a statement from the American Association for Cancer Research (AACR).

Multiple pulmonary nodules are commonly detected on CT lung cancer screening exams; approximately half of patients found to have a nodule in a previous lung cancer screening trial had multiple nodules, according to co-first author Dr. Kezhong Chen of Peking University People’s Hospital in China.

Screenshot of the web-based PKU-M model

Screenshot of the web-based PKU-M model’s analysis of a 64-year-old male patient with a family history of lung cancer, a 50-pack-year smoking history, and an elevated carcinoembryonic antigen value. An incidental thoracic CT scan discovered two nodules scattered bilaterally in the lung. The nodule located in the right lower lobe was 21 mm and presented spiculation, lobulation, the pleural retraction sign, and the solid type. The predicted probability of cancer by the PKU-M model was 87.4%. After surgical resection, pathologic examination confirmed squamous cell carcinoma. Image and caption courtesy of Clinical Cancer Research.

“Current guidelines recommend the use of clinical models that incorporate nodule and sociodemographic features to estimate the probability of cancer prior to surgical treatment, and while there are several tools for patients that present with a single nodule, no such tool currently exists for patients with multiple nodules, representing an urgent medical need,” Chen said.

The researchers created PKU-M to address this challenge. The machine-learning algorithm was trained using radiographic nodule characteristics and sociodemographic variables from 520 patients at Peking University People’s Hospital who had a total of 1,739 pulmonary nodules.

After producing an area under the curve (AUC) of 0.91 in the training cohort, PKU-M then underwent external validation on a cohort of 220 patients with 583 nodules who had received surgery at six other hospitals in China and Korea between January 2016 and December 2018. The algorithm achieved an AUC of 0.89, compared with an AUC range of 0.68 to 0.81 for four prior logistic regression-based lung cancer risk prediction models.

Next, they prospectively evaluated PKU-M on a separate cohort of 78 patients with 200 nodules in comparison with a chest radiologist with five years of experience, three thoracic surgeons with experience ranging from three to 10 years, and a previously developed AI model — RX. The patients were treated surgically at four independent hospitals in China between January 2019 and March 2019.

Performance of PKU-M model for predicting lung nodule malignancy
Thoracic surgeons Radiologist RX model PKU-M algorithm
AUC 0.73-0.79 0.75 0.76 0.87

PKU-M yielded 14.3% higher sensitivity and 7.8% greater specificity than the clinicians, according to the researchers.

“This tool can quickly generate an objective diagnosis and can aid in clinical decision-making,” said Dr. Jun Wang of Peking University People’s Hospital.

A web-based version of the model has been created, enabling clinicians to input several clinical and radiological characteristics and receive a risk prediction. As their algorithm was developed and tested only on Asian patients, the researchers acknowledged that it may not be generalizable to Western populations, however.