UCLA

Idiopathic pulmonary fibrosis (IPF) is one type of interstitial lung disease (ILD) of unknown causes. High-resolution computed tomography (HRCT) scans play a crucial role in distinguishing IPF from non-IPF among subjects with ILD. This radiological evaluation is an important yet difficult task. In clinical practice, making a correct and reliable IPF diagnosis is critical to ensure patients with different causes of pulmonary fibrosis be treated appropriately and patients with IPF be assessed for novel therapies and lung transplantation. Therefore, this dissertation aims to build an automated IPF diagnosis system for ILD subjects, using volumetric and non-contrast chest HRCT scans.

Supervised learning methods are a type of machine learning methods that require labels of training samples as ground truth to learn a mapping function between input and output labels. Depending on the type of labels provided, if only coarse-level labels are available, rather than fine-scale labels, this task is called a weakly supervised task. For our example, acquiring fine-scale information of certain radiological patterns can be helpful for building the diagnostic system, but fine-scale labels are expensive to obtain. On the other hand, coarse-level labels are usually easier to acquire, such as CT scan-level information. Since we only have labels at a CT scan level, our problem is a weakly supervised task. To tackle this challenge, this dissertation leverages domain knowledge acquired from previous studies, including IPF progression and quantification information, to provide more efficient, reliable, and explainable diagnostic support.

In project I, we used 2D deep learning models with IPF progression information and optimal design criterion to weigh HRCT samples differently. In project II, 3D deep learning models with multi-scale attention models were used with IPF quantification maps to achieve good model accuracy and explainability. Furthermore, we evaluated the robustness of these developed models under a different set of HRCT parameters, using paired HRCT scans. These proposed methodologies in projects I and II can be applied to other weakly supervised tasks, where domain knowledge is available. The method used in the robustness tests can be applied to evaluate model performance if paired medical images are available.