%0 Thesis %9 Doctoral %A Zhao, An %B Computer Science %D 2024 %F discovery:10192608 %I UCL (University College London) %P 234 %T Machine and Deep Learning Methods for Prognosis Prediction of Idiopathic Pulmonary Fibrosis %U https://discovery.ucl.ac.uk/id/eprint/10192608/ %X Idiopathic pulmonary fibrosis (IPF) is a chronic, progressive, life-threatening interstitial pneumonia of unknown etiology, characterized by progressive lung fibrosis, deterioration in respiratory symptoms, lung function decline, and early mortality. IPF has a highly heterogeneous disease course, making accurate prognostication quite challenging. Despite many efforts in this field, to date, we have limited prognostic biomarkers, and research in this area hasn’t built widely accepted models for disease progression prediction and outcome prediction that can be reliably used in the clinical practice of IPF patients. High-resolution computed tomography (HRCT) scanning provides rich information about the nature of the disease and plays an important role in IPF prognosis. Advanced machine learning methods have created new opportunities for mining underlying prognostic information in HRCT scans. In this thesis, I explore machine learning models in various applications of HRCT for improving prognostication and disease progression modelling of IPF. My contributions are threefold. First, to better monitor the heterogeneous IPF patients, I use a threshold method and a machine learning method to subtype patients into different subgroups and then identify the optimal functional indicators as mortality surrogates for different subtypes. Second, I propose a method for discovering patch-level imaging patterns that I then use to predict mortality risk and identify prognostic biomarkers. Moreover, by comparing the high-risk imaging patterns extracted by the proposed model with existing imaging patterns utilized in clinical practice, I identify a novel biomarker that may help clinicians improve the risk stratification of IPF patients. Third, to involve longitudinal data that captures more information about disease evolution, I propose a model for disease progression modelling from irregularly sampled longitudinal 3D imaging data. The model can generate synthetic volumetric CT scans at any time point given sparse observations and does not rely on strong assumptions about disease evolution dynamics. %Z Copyright © The Author 2024. Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) Licence (https://creativecommons.org/licenses/by-nc/4.0/).