Improving clinical disease subtyping and future events prediction through a chest CT‐based deep learning approach

Sumedha Singla, Mingming Gong, Craig Riley, Frank Sciurba, Kayhan Batmanghelich
2021-03-01
Abstract:Purpose To develop and evaluate a deep learning (DL) approach to extract rich information from high‐resolution computed tomography (HRCT) of patients with chronic obstructive pulmonary disease (COPD). Methods We develop a DL‐based model to learn a compact representation of a subject, which is predictive of COPD physiologic severity and other outcomes. Our DL model learned: (a) to extract informative regional image features from HRCT; (b) to adaptively weight these features and form an aggregate patient representation; and finally, (c) to predict several COPD outcomes. The adaptive weights correspond to the regional lung contribution to the disease. We evaluate the model on 10 300 participants from the COPDGene cohort. Results Our model was strongly predictive of spirometric obstruction ( =  0.67) and grouped 65.4% of subjects correctly and 89.1% within one stage of their GOLD severity stage …
What problem does this paper attempt to address?