Deep learning-enabled objective image quality assessment in medical imaging

It has been advocated to use objective, or task-based, measures of image quality (IQ) in the assessment and optimization of medical imaging systems. Unlike traditional physical measures of IQ, task-based measures of IQ quantify the ability of an observer to perform a specific task such as detection or estimation of a signal. Among all numerical observers (NOs), the Bayesian ideal observer (IO) achieves the performance of the optimal decision maker acting on given measured data. While such IO analyses are known conceptually, they have generally remained analytically intractable to widely implemented in both image space and data space. The applications of objective IQ assessment (OIQA) have become even more critical due to an unprecedented revolution in medical image restoration prompted by deep learning (DL). New and innovative DL-based medical image restoration methods (DLIR) and reconstruction methods are being developed at a breakneck pace to address impactful medical imaging applications. However, the overwhelming majority of DLIRs reported to date have only been assessed by use of task-agnostic metrics from the field of computer vision instead of metrics that convey the potential utility of the restored images for specified medical imaging tasks. Considering the fact that medical images are often acquired for specific purposes, this has resulted in a large and continuously growing collection of DLIRs whose suitability and reliability for clinical applications remain largely unknown. In the first part of this dissertation, advanced machine learning-enabled IO approximation methods were developed for applications in both image space and data space. Specifically, we first developed machine learning-enabled IO approximation methods and sub-ideal observers for general detection-estimation tasks. Computer-simulation studies were conducted to validate the proposed methods. Next, the performance of the IO acting on the raw measured data to perform a diagnostically relevant task was estimated to establish task performance bounds for image reconstruction. This new capacity can be employed to identify situations where tomographic measurement data will not permit the reconstruction of a diagnostically useful image—no matter how advanced the reconstruction method is or plausible the reconstructed images appear. The second part of this dissertation focuses on evaluation studies for both conventional and task-informed DL-based denoising methods to demonstrate the application of the developed OIQA techniques. Specifically, comprehensive and systematic studies were designed to assess the impact of the choice and depth of network architecture on the utility of the restored images for different clinically relevant diagnostic tasks. This study will result in fundamentally new and medically relevant insights into DLIRs that exist today and reveal opportunities for their improvements. Finally, fundamental issues regarding learning-based denoising methods that incorporate task-relevant information into the training procedure were investigated. This includes understanding the impact of denoising on objective measures of IQ when the specified task at inference time is different from that employed for model training, a phenomenon we refer to as ``task-shift". Our studies can provide guidance for the deployment of these methods in a clinically-relevant scenario.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Bioengineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2024

Author dc:creator

• Li, Kaiyan

Contributors dc:contributor

• Li, Hua
• Anastasio, Mark
• Insana, Michael
• Lam, Fan

• Task-based Image Quality Assessment
• Numerical Observer
• Ideal Observer
• Medical Imaging
• Deep Learning

Statement dc:rights

• Copyright 2024 Kaiyan Li

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/127133

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F127133