FibroRegNet: A Regression Framework for the Pulmonary Fibrosis Prognosis Prediction Using a Convolutional Spatial Transformer Network

Abstract

Predicting the growth of idiopathic pulmonary fibrosis (IPF) is crucial for effectively treating patients affected by the disease. While the Forced Vital Capacity (FVC) serves as one of the indicators of lung functionality, accurately determining its decline solely based on previous FVC values presents a significant obstacle. We propose the utilization of a multimodal system called FibroRegNet, which capitalizes on the recent achievements in cross-model learning across general domains. FibroRegNet is designed to acquire knowledge through the regression function which maps the multimodal inputs, including CT scan and demographic information, to the coefficients of the quadratic polynomial ridge regression of FVC as outputs. FibroRegNet estimates the lung volume from a fraction of CT slices, encodes the demographic information, and combines these features with the convolutional features, from selected CT slices, that are learned through convolutional spatial transformer modules in three identical parallel streams. Trained on a publicly available database, FibroRegNet has shown significant improvement in the results compared to the related past works with a modified Laplace log-likelihood score of −6.64. Furthermore, we believe that this network has the potential to provide advantages in research domains related to the development of networks aimed at enhancing the predictive precision of IPF.

Keywords

• CNN
• fibrosis prognosis
• IPF
• lung CT scans
• regression
• spatial transformer