Deep Learning and Subtyping of Post-COVID-19 Lung Progression Phenotypes

PROJECT SUMMARY Patients who recover from the novel coronavirus disease 2019 (COVID-19) may experience a range of long- term health consequences. Since the lung is the primary site of viral infection, pulmonary sequelae may present persistently in COVID-19 survivors. Thus, clinical assessment of COVID-19 survivors in conjunction with chest X-ray (CXR) and computed tomography (CT) is recommended. CXR is more accessible, whereas CT provides more detailed information. Our long-term goal is to develop an integrated deep learning model that can assess lung images to assist with the management and treatment of long-term sequelae of post-COVID-19 subjects. The primary objective of the proposed research is to advance contrastive self-supervised learning models that take advantage of the accessibility of CXR scanners and the accuracy of CT images, identify the subtypes in patients with post-COVID-19, and characterize clinical, imaging and mechanistic biomarkers within subtypes. Our central hypothesis is that post-COVID-19 subtypes exist and they are characterized by distinct progression phenotypes. To test this hypothesis and achieve the primary objective, we will perform the following four specific aims. In Aim 1, we will advance contrastive learning methods to handle large-scale images with low training costs, and fine-tune the classifier and the encoder network on large-scale CXR images to detect post-COVID- 19 subjects. In Aim 2, we will advance contrastive learning methods that learn from CT images acquired at different volumes and different times to differentiate post-COVID-19 subjects from other cohorts and identify subtypes. In Aim 3, we will apply computational fluid and particle dynamics techniques to derive mechanistic biomarkers to explain the associations between clinical and imaging biomarkers in post-COVID-19 subtypes. In Aim 4, we will conduct a human subject study that examines post-COVID-19 subjects at 36-48 months after initial follow-up visits to assess the progression features of their clinical and imaging biomarkers. In summary, we will advance contrastive self-supervised learning algorithms based on CXR and CT images, respectively, for accessibility (Aim 1) and accuracy (Aim 2). We will generate in silico data for feature interpretability (Aim 3) and gather in vivo data for model training and validation (Aim 4). The pre-trained model from Aim 2 will be fine-tuned via transfer learning to input CXR images that are classified as post-COVID-19 by the model from Aim 1. An integrated deep learning model based on the two models from Aim 1 and 2 will take CXR images as inputs to provide CT-based detailed phenotypic information together with mechanistically and clinically meaningful interpretation. If successful, our study will not only advance contrastive learning algorithms, but also elucidate the pulmonary sequelae of post-COVID-19 patients in subtypes and associated clinical, imaging and mechanistic biomarkers. The ability to identify progression subtypes and associated phenotypic biomarkers will have a positive impact on the management and treatment of patients with post-COVID-19.