Presented by by Dr. Fenghai Duan, PhD, Associate Professor of Biostatistics

Lung cancer is the leading cause of cancer death worldwide each year. Non-invasive medical image technologies are becoming routine in screening high-risk populations for lung cancer patients. Unlike traditional radiological imaging analysis, which is manually interpreted by radiologists, the rapid progress of computational methods and artificial intelligence is leading to the extensive implementation of radiomics and/or deep learning in medical image analysis. Specifically, radiomics involves the high-throughput extraction and analysis of quantitative features from advanced medical images, with assistance from compute science, to provide a comprehensive quantification of tumor phenotype in cancer patients. Deep learning generally refers to the application of deep neural networks, such as CNNs, to process and analyze medical images. In this study, we analyzed clinical and imaging data from the National Lung Screening Trial to construct three risk models for discriminating lung nodule malignancy. These included a model built from semantic features, a radiomic model built from radiomic features, and a deep learning model developed using transfer learning. The performance of these three models was thoroughly assessed and compared.