Automated Pulmonary Function Measurements from Preoperative CT Scans with Deep Learning

Lung resections are the most effective treatment option for early stage lung cancer. Clinicians determine whether a patient is operable and the extent a lung can be resected based in part on the patient's pulmonary function parameters. In this study, we investigate the feasibility of generating forced expiratory volume in 1 second (FEV1) and forced vital capacity (FVC) values from preoperative chest computed tomography (CT) scans. Our study population includes 546 individuals who had lung cancer surgery at an oncology specialty clinic between 2009 and 2015. All CT studies and pulmonary function tests (PFTs) were collected within 90 days before a subject's operation. We measure pulmonary function with convolutional neural network and recurrent neural network models, extracting image embeddings from axial CT slices with a ResNet-50 network and generating FEV1 and FVC measurements using a bidirectional long short-term memory regressor. We show that combining feature vectors extracted from mediastinal and lung Hounsfield unit windows and taking a multi-label regression approach improves performance over training with embeddings from only one window or single-task networks trained to measure only FEV1 or FVC values. Our work generates PFT measurements end-to-end and is trained with only computed tomography scans and pulmonary function labels with no manual slice selection, bounding boxes, or segmentation masks.