Deep learning for assessing pulmonary parenchyma disease in smoking-related injury

Resumen

Tobacco smoking is one of the health risk factors most associated with population morbidity and mortality causing a variety of lung diseases that are responsible for the death of more than 8 million people each year. The inhalation of cigarette smoke can affect the lung in diverse ways and is directly associated with the development of lung parenchyma disorders including emphysema, characterized by airspace dilation and alveolar destruction, and interstitial or fibrotic changes caused by the inflammation and scarring of pulmonary parenchyma. These parenchymal injuries are often an irreversible process specially when they are in an advanced stage. High resolution computed tomography (HRCT) plays an essential role in the diagnosis and assessment of disease prognosis and progression of lung diseases. In this thesis, we developed deep learning approaches to characterize parenchymal injury patterns on HRCT, which may facilitate disease understanding, serve as a strong basis to assess disease prognosis and progression, and facilitate automated diagnosis of lung diseases at early stages. In this Ph.D. thesis, we described a novel optimally cost-effective convolutional neural network (CNN) to characterize emphysema subtypes which was subsequently extended by developing an ensemble of multiple CNN covering the lesions' multiscale representation designed to identify interstitial lung abnormalities that are found as early sign of fibrotic disorders in the lung parenchyma. Moreover, we proposed an innovative CNN segmentation architecture to identify paraseptal emphysema lesions, a particular subtype of emphysema with a unique notion of location, that leverages and exploits 3D contextual information while producing 2D annotations alleviating the need of having 3D labels which carries a high burden from experts. Finally, we designed a novel methodology that learns the spatial interdependence of parenchymal injuries patterns using a segmentation approach from an initial labelling based on local methods. This new paradigm relieves the need for delimited annotations to perform segmentation tasks. Developed approaches have demonstrated that are able to identify subtype injury patterns which are related to different clinical outcomes. Therefore, these methods allow to advance in the understanding of both lung diseases and the clinical implications of the different parenchymal patterns.