Deep Learning in Chest Radiography: From Report Labeling to Image Classification
Access status:
Open Access
Type
ThesisThesis type
Doctor of PhilosophyAuthor/s
Monshi, Maram Mahmoud AAbstract
Chest X-ray (CXR) is the most common examination performed by a radiologist. Through CXR, radiologists must correctly and immediately diagnose a patient’s thorax to avoid the progression of life-threatening diseases. Not only are certified radiologists hard to find but also stress, ...
See moreChest X-ray (CXR) is the most common examination performed by a radiologist. Through CXR, radiologists must correctly and immediately diagnose a patient’s thorax to avoid the progression of life-threatening diseases. Not only are certified radiologists hard to find but also stress, fatigue, and lack of experience all contribute to the quality of an examination. As a result, providing a technique to aid radiologists in reading CXRs and a tool to help bridge the gap for communities without adequate access to radiological services would yield a huge advantage for patients and patient care. This thesis considers one essential task, CXR image classification, with Deep Learning (DL) technologies from the following three aspects: understanding the intersection of CXR interpretation and DL; extracting multiple image labels from radiology reports to facilitate the training of DL classifiers; and developing CXR classifiers using DL. First, we explain the core concepts and categorize the existing data and literature for researchers entering this field for ease of reference. Using CXRs and DL for medical image diagnosis is a relatively recent field of study because large, publicly available CXR datasets have not been around for very long. Second, we contribute to labeling large datasets with multi-label image annotations extracted from CXR reports. We describe the development of a DL-based report labeler named CXRlabeler, focusing on inductive sequential transfer learning. Lastly, we explain the design of three novel Convolutional Neural Network (CNN) classifiers, i.e., MultiViewModel, Xclassifier, and CovidXrayNet, for binary image classification, multi-label image classification, and multi-class image classification, respectively. This dissertation showcases significant progress in the field of automated CXR interpretation using DL; all source code used is publicly available. It provides methods and insights that can be applied to other medical image interpretation tasks.
See less
See moreChest X-ray (CXR) is the most common examination performed by a radiologist. Through CXR, radiologists must correctly and immediately diagnose a patient’s thorax to avoid the progression of life-threatening diseases. Not only are certified radiologists hard to find but also stress, fatigue, and lack of experience all contribute to the quality of an examination. As a result, providing a technique to aid radiologists in reading CXRs and a tool to help bridge the gap for communities without adequate access to radiological services would yield a huge advantage for patients and patient care. This thesis considers one essential task, CXR image classification, with Deep Learning (DL) technologies from the following three aspects: understanding the intersection of CXR interpretation and DL; extracting multiple image labels from radiology reports to facilitate the training of DL classifiers; and developing CXR classifiers using DL. First, we explain the core concepts and categorize the existing data and literature for researchers entering this field for ease of reference. Using CXRs and DL for medical image diagnosis is a relatively recent field of study because large, publicly available CXR datasets have not been around for very long. Second, we contribute to labeling large datasets with multi-label image annotations extracted from CXR reports. We describe the development of a DL-based report labeler named CXRlabeler, focusing on inductive sequential transfer learning. Lastly, we explain the design of three novel Convolutional Neural Network (CNN) classifiers, i.e., MultiViewModel, Xclassifier, and CovidXrayNet, for binary image classification, multi-label image classification, and multi-class image classification, respectively. This dissertation showcases significant progress in the field of automated CXR interpretation using DL; all source code used is publicly available. It provides methods and insights that can be applied to other medical image interpretation tasks.
See less
Date
2022Rights statement
The author retains copyright of this thesis. It may only be used for the purposes of research and study. It must not be used for any other purposes and may not be transmitted or shared with others without prior permission.Faculty/School
Faculty of Engineering, School of Computer ScienceAwarding institution
The University of SydneyShare