Persistent Homology and Machine Learning of Radiological Images for Characterization and Prediction of Interstitial Lung Disease
PI:
Jason Cory Brunson, PhD
Bruno Hochhegger, MD, PhD
Hubert Wagner, PhD
Project Summary
Interstitial lung disease (ILD) is caused by several disorders and carries significant risk of irreversible and life-threatening scarring called pulmonary fibrosis. Fibrosis is difficult to predict or detect, and efforts continue to discover early markers. Conventional medical image analysis approaches, including radiological feature sets and convolutional neural networks (CNNs), struggle to robustly capture such patterns. A promising alternative comes from topological data analysis (TDA), which has shown success on related biomedical imaging tasks when combined with CNNs and other machine learning (ML). The main limitation of TDA in this setting has been its computational cost. This project addresses these limitations by building on GPU-accelerated software for TDA of large 3D data, recently developed by PI Wagner and collaborators. The goal is to develop a custom, scalable analysis pipeline integrated with modern ML tools that can leverage the power of HiPerGator, with which to achieve accurate, automated assessment of ILD-related lung damage. We aim to develop a newly practical approach to automatic detection and risk assessment for ILD from 3D computational tomography (CT) images. The project team consists of three PIs with complementary expertise in ILD, thoracic radiology, medical image analysis, machine learning, and topological data analysis.