Automatic analysis of chest CT in systemic sclerosis using deep learning

Doctoral Thesis

The aim of this thesis is to develop automatic methods for quantifying the severity of SSc disease from CT images through direct and indirect routes. The indirect route involves lung, lobe, and vessel segmentation (Chapter 2) and PFT estimation from segmented vessels (Chapter 5). The direct route focuses on directly estimating PFT (Chapter 4) and scoring ILD from CT (Chapter 3). Chapter 2 introduces a deep-learning network for lobe segmentation using a multi-task semi-supervised model and an alternating training strategy, evaluated on an external CT dataset. A Python package for calculating segmentation metrics is developed (see Chapter 8 Supplementary material). Chapter 3 presents a deep learning framework for automating SSc-ILD scoring, using a cascade of two neural networks to select craniocaudal positions and estimate pattern ratios, with synthetic data augmentation and a heat map method for output explanation. Chapter 4 proposes a deep-learning...Show moreThe aim of this thesis is to develop automatic methods for quantifying the severity of SSc disease from CT images through direct and indirect routes. The indirect route involves lung, lobe, and vessel segmentation (Chapter 2) and PFT estimation from segmented vessels (Chapter 5). The direct route focuses on directly estimating PFT (Chapter 4) and scoring ILD from CT (Chapter 3). Chapter 2 introduces a deep-learning network for lobe segmentation using a multi-task semi-supervised model and an alternating training strategy, evaluated on an external CT dataset. A Python package for calculating segmentation metrics is developed (see Chapter 8 Supplementary material). Chapter 3 presents a deep learning framework for automating SSc-ILD scoring, using a cascade of two neural networks to select craniocaudal positions and estimate pattern ratios, with synthetic data augmentation and a heat map method for output explanation. Chapter 4 proposes a deep-learning framework for automatic PFT estimation from CT scans, exploring the influence of segmented lungs and vessels, and introducing regression attention maps (RAM) to highlight contributing regions. Chapter 5 extends Chapter 4's work by enhancing PFT estimation performance with point cloud (PNN-Vessel) and graph neural networks (GNN-Vessel) based on vessel centerlines, and combining different networks for optimal results.Show less

Lobe segmentation

Artificial intelligence

Pulmonary function

Systemic sclerosis