Encompassing with fields on engineering and medical image quality, this dissertation proposes a novel framework for diagnostic performance evaluation based on objective image-quality assessment, an important step in the development of new imaging devices, acquisitions, or image-processing techniques being used for clinicians and researchers. The objective of this dissertation is to develop computational modeling tools that allow comprehensive evaluation of task-based assessment including clinical interpretation of images regardless of image dimensionality.
Because of advances in the development of medical imaging devices, several techniques have improved image quality where the format domain of the outcome images becomes multidimensional (e.g., 3D+time or 4D). To evaluate the performance of new imaging devices or to optimize various design parameters and algorithms, the quality measurement should be performed using an appropriate image-quality figure-of-merit (FOM). Classical FOM such as bias and variance, or mean-square error, have been broadly used in the past. Unfortunately, they do not reflect the fact that the average performance of the principal agent in medical decision-making is frequently a human observer, nor are they aware of the specific diagnostic task.
The standard goal for image quality assessment is a task-based approach in which one evaluates human observer performance of a specified diagnostic task (e.g. detection of the presence of lesions). However, having a human observer performs the tasks is costly and time-consuming. To facilitate practical task-based assessment of image quality, a numerical observer is required as a surrogate for human observers. Previously, numerical observers for the detection task have been studied both in research and industry; however, little research effort has been devoted toward development of one utilized for multidimensional imaging studies (e.g., 4D). Limiting the numerical observer tools that accommodate all information embedded in a series of images, the performance assessment of a particular new technique that generates multidimensional data is complex and limited. Consequently, key questions remain unanswered about how much the image quality improved using these new multidimensional images on a specific clinical task.
To address this gap, this dissertation proposes a new numerical-observer methodology to assess the improvement achieved from newly developed imaging technologies. This numerical observer approach can be generalized to exploit pertinent statistical information in multidimensional images and accurately predict the performance of a human observer over the complexity of the image domains. Part I of this dissertation aims to develop a numerical observer that accommodates multidimensional images to process correlated signal components and appropriately incorporate them into an absolute FOM. Part II of this dissertation aims to apply the model developed in Part I to selected clinical applications with multidimensional images including: 1) respiratory-gated positron emission tomography (PET) in lung cancer (3D+t), 2) kinetic parametric PET in head-and-neck cancer (3D+k), and 3) spectral computed tomography (CT) in atherosclerotic plaque (3D+e).
The author compares the task-based performance of the proposed approach to that of conventional methods, evaluated based on a broadly-used signal-known-exactly /background-known-exactly paradigm, which is in the context of the specified properties of a target object (e.g., a lesion) on highly realistic and clinical backgrounds. A realistic target object is generated with specific properties and applied to a set of images to create pathological scenarios for the performance evaluation, e.g., lesions in the lungs or plaques in the artery. The regions of interest (ROIs) of the target objects are formed over an ensemble of data measurements under identical conditions and evaluated for the inclusion of useful information from different complex domains (i.e., 3D+t, 3D+k, 3D+e). This work provides an image-quality assessment metric with no dimensional limitation that could help substantially improve assessment of performance achieved from new developments in imaging that make use of high dimensional data.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/D8Z60NB4 |
| Date | January 2015 |
| Creators | Lorsakul, Auranuch |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.0023 seconds