Diagnostic Classification of Lung CT Images Using Deep 3D Multi-Scale Convolutional Neural Network

Lung cancer is the second most common cancer diagnosis and the leading cause of cancer mortality worldwide. Accurate and early diagnosis of cancerous lungs is of paramount importance to provide right treatment plan at the right time, avoiding unnecessary medical costs due to biopsy, surgery, and prolonged hospitalization. In this contribution, we designed and developed a deep 3D multi-scale convolutional neural network to classify cancerous versus non-cancerous lesions in lung CT images. The proposed deep learning model was trained, built, and evaluated using a large-scale dataset obtained from publicly available data at the Data Science Bowl 2017 plus privately held medical imaging data from Marshfield Clinic's electronic health records, and it can be employed on smaller datasets with the use of transfer learning strategy. The results reveal that the current computational method aims to provide a reliable data-driven computer-aided diagnosis (CADx) system which reduces the false positive rate, and assists radiologists and physicians to get closer to more accurate and quicker diagnosis.