Investigation of Data Modeling Strategies for Quantification of CT Image Quality

Abstract

Computed Tomography (CT) is a widely used medical imaging technology that plays a crucial role in pathology diagnosis and treatment management. The growing use of CT imaging raises the risk of undue radiation exposure, particularly for patients imaged multiple times. Thus, there is a need to maintain radiation exposure As Low As Reasonably Achievable (ALARA) while ensuring diagnostic quality images. In this study, we investigate data modeling strategies to quantify CT image quality (IQ), in order to guide protocol selection and attain ALARA. Specifically, we present a novel, Windowed Fourier-domain Distance Metric (WFDM) that is used to select regions-of-interest (ROI) by their degree of spatial variability. By selecting regions of low variation (ROI-LV), an estimate of the noise in that region can be made. CT IQ is defined as the inverse of this additive noise. Against the phantom CT images, the WFDM model is shown to correlate strongly to image noise (r > 0.76 (p 0.001)). As a CT IQ classifier, this model is comparatively analyzed against a fixed-size ROI (baseline) model and a Convolutional Neural Network (CNN), using phantom and patient CT images. The WFDM model and the CNN are shown to classify the phantom images accurately, with a mean accuracy of αWFDM ≤ 100%, and αCNN = 93.8%, respectively. The baseline model manages a mean accuracy of αB = 73.6% on the same phantom images. With the patient CT images, the baseline and WFDM accuracies drop to αB ≤ 49.5% and αWFDM ≤ 66.1%, respectively. The CNN, however, performs at 100% accuracy when tested with images from the same CT stack as the training set, but below 1.9% otherwise. This indicates the CNN focus on structural rather than textural features. Finally, the WFDM model is used to predict high/low trends in 84 pairs of patient CT images. These trends are set against the trends in x-ray flux at the time of acquisition, CTDIvol, which, for the same patient, directly correspond to CT IQ. The total percentage of image pairs with inverse trends is defined as the total percent error, which is found to be 30.95% and 21.43% for the baseline and WFDM models, respectively. However, this error drops to 0% for CTDIvol changes of at least 40.0% for the baseline model and 27.5% for the WFDM model, respectively. Thus, for every patient that has been previously imaged, the WFDM model can be used to predict optimal parameters for adequate CT image acquisition. Future work will investigate the impact of WFDM parameters, such as the window sizes and transformation technique, on CT image quality assessment. In addition, the WFDM model can also be applied to pre-process CT images followed by CNN data models for CT image texture identification.