Patient stratification in lung cancer screening using SmileGAN

While lung cancer remains the deadliest cancer worldwide, completed lung cancer screening trials have shown a significant decrease in lung cancer mortality, but at the price of a high overdiagnosis rate. This work aimed to assess the potential of SmileGAN, a generative adversarial network for semi-supervised clustering, in patient stratification for lung cancer screening. Since the completed lung cancer screening trials showed that the major causloss; death in the screening cohort were cardiovascular disease, lung cancer and respiratory illnesses, we wanted to ascertain whether SmileGAN was able to cluster these three cohorts of patients. In each of the three test scenarios, namely synthetic, semi-synthetic and ‘real-world’, SmileGAN clustering performance surpassed the K-means one, reaching an accuracy of 0.58 and 0.97 respectively in the cases where 3 or 2 clusters where considered.