Author ORCID Identifier
Date of Graduation
Doctor of Philosophy (PhD)
Laurence E. Court
Clifton D. Fuller
Rebecca M. Howell
Rick R. Layman
R. Jason Stafford
The purpose of this work was to determine if prediction models using quantitative imaging measures in head and neck squamous cell carcinoma (HNSCC) patients could be improved when noise due to imaging was reduced. This was investigated separately for salivary gland function using dynamic contrast enhanced magnetic resonance imaging (DCE-MRI), overall survival using computed tomography (CT)-based radiomics, and overall survival using positron emission tomography (PET)-based radiomics. From DCE-MRI, where T1-weighted images are serially acquired after injection of contrast, quantitative measures of diffusion can be obtained from the series of images. Radiomics is the study of the relationship of voxels to one another providing measures of texture from the area of interest. Quantitative information obtained from imaging could help in radiation treatment planning by providing quantifiable spatial information with computational models for assigning dose to regions to improve patient outcome, both survival and quality of life. By reducing the noise within the quantitative data, the prediction accuracy could improve to move this type of work closer to clinical practice.
For each imaging modality sources of noise that could impact the patient analysis were identified, quantified, and if possible minimized during the patient analysis. In MRI, a large potential source of uncertainty was the image registration. To evaluate this, both physical and synthetic phantoms were used, which showed that registration of MR images was high, with all root mean square errors below 3 mm. Then, 15 HNSCC patients with pre-, mid-, and post-treatment DCE-MRI scans were evaluated. However, differences in algorithm output were found to be a large source of noise as different algorithms could not consistently rank patients as above or below the median for quantitative metrics from DCE-MRI. Therefore, further analysis using this modality was not pursued.
In CT, a large potential source of noise that could impact patient analysis was the inter-scanner variability. To investigate this a controlled protocol was designed and used to image, along with the local head and chest protocols, a radiomics phantom on 100 CT scanners. This demonstrated that the inter-scanner variability could be reduced by over 50% using a controlled protocol compared to local protocols. Additionally, it was shown that the reconstruction parameters impact feature values while most acquisition parameters do not, therefore, most of this benefit can be achieved using a radiomics reconstruction with no additional dose to the patient. Then to evaluate this impact in patient studies, 726 HNSCC patients with CT images were used to create and test a Cox proportional hazards model for overall survival. Those patients with the same imaging protocol were subset and a new Cox proportional hazards model was created and tested in order to determine if the reduction in noise due to controlling the imaging protocol translated into improved prediction. However, noise between patient populations from different institutions was shown to be larger than the reduction in noise due to a controlled imaging protocol.
In PET, a large potential source of noise that could impact patient analysis was the imaging protocol. A phantom scanned on three different scanners and vendors demonstrated that on a single vendor, imaging parameter choices did not affect radiomics feature values, but inter-scanner variances could be large. Then, 686 HNSCC patients with PET images were used to create and test a Cox proportional hazards model for overall survival. Those patients with the same imaging protocol were subset and a new Cox proportional hazards model was created and tested in order to determine if the reduction in noise due to controlling the imaging protocol on a vendor translated into improved prediction. However, no predictive radiomics signature could be determined for any subset of the patient cohort that resulted in significant stratification of patients into high and low risk.
This study demonstrated that the imaging variability could be quantified and controlled for in each modality. However, for each modality there were larger sources of noise identified that did not allow for improvement in prediction modeling of salivary gland function or overall survival using quantitative imaging metrics for MRI, CT, or PET.
computed tomography, positron emission tomography, magnetic resonance imaging, radiomics, survival prediction, quantitative imaging, uncertainty
Available for download on Thursday, April 09, 2020