Predicting ARDS trajectory with deep generative models of CT images

PROJECT SUMMARY Acute respiratory distress syndrome (ARDS) is characterized by inflammatory lung injury, heterogeneous pulmonary inflation, and hypoxemia. It requires mechanical ventilation to prevent immediate death. Tuning ventilation settings to minimize harmful lung stretch, and ventilating patients in the prone position mitigate lung injury and reduce mortality. However, responses to these interventions are not uniform and clinical progression is poorly predictable. Clinical studies suggested that regional patterns of lung aeration (e.g., radiographic consolidation vs. disseminated air loss) may be associated with distinct outcomes and different responses to ventilatory strategies. Topological assessment via computed tomography (CT) could predict treatment responses and facilitate personalized ARDS care, but this application of CT is hindered by subjective interpretation and by the burden of image processing. Artificial intelligence (AI) algorithms, particularly deep learning, are revolutionizing image recognition and processing tasks. Leveraging the latest AI developments in image synthesis, the goal of this project is to develop, train, and test deep generative models using CT images to objectively and rapidly predict lung injury progression and treatment responses. We will develop generative adversarial networks (GANs) to synthesize: a) CT images that emulate responses to interventions, and b) CT images that predict future injury morphology. The models will be trained using over 2000 paired CT images from different conditions, such as respiratory treatments and injury stages, previously obtained from pigs with experimental ARDS. Moreover, we will create an image augmentation network with GANs to address the potential scarcity of clinical datasets and possibly enhance prediction performance. Lastly, we will evaluate the generative models in human ARDS cases. We will begin by retraining them to predict treatment responses using existing clinical imaging datasets. Subsequently, we will predict injury progression with prospectively collected longitudinal CT scans. We will also integrate clinical parameters, physiological measurements, and inflammatory biomarkers to refine and enhance the generative models' ability to predict ARDS progression and patient outcomes. This project has several innovative aspects. We will be the first group applying deep generative networks to synthesize lung CT images in ARDS, using these to predict treatment responses and injury progression. If successful, the integration of the developed techniques in personalized clinical decision-making could significantly impact ARDS. We hypothesize that the multi-layer neural networks will capture the hidden characteristics of the ARDS lung after learning thousands of pairs of CT images obtained in various conditions, such as different interventions and stages of injury. We expect this AI approach will eventually enable quantitative imaging for prognostication and treatment selection.