Pulmonary fibrosis progression prediction based on ResNet50 architecture

Abstract

Pulmonary fibrosis (PF) is a chronic lung disease which causes permanent scarring of lung tissue over time. Current treatment options focus on slowing down the scarring process and increasing quality of life for the patient. However, due to the variable rate of progression between patients, and the potential lack of visual signs, there is still a need to develop new methods of assessing CT scans in order to obtain more accurate prediction results. This paper evaluates a convolutional neural network (CNN)-based method to predict the disease progression using the path of Forced Vital Capacity (FVC) decline, with the hopes that it could better aid doctors in managing the care of patients suffering from PF. The proposed model uses a ResNet50 backbone to predict a linear rate of FVC decline from inputs consisting of a baseline CT scan, and tabular patient data. The predicted gradient will then be used to calculate the predicted FVC value on a weekly basis. The model is trained and evaluated on a dataset obtained from the OSIC Pulmonary Fibrosis Progression challenge, which also served as the benchmark for the evaluation of the proposed model. The proposed model was not able to obtain any improvement over the state-of-the-art solutions, and changes to the current models are suggested for future improvement.