Hybrid Reference Datasets for Quantitative Computed Tomography Characterization and Conformance
X-ray computed tomography (CT) imaging is the second most commonly used clinical imaging modality with an estimated 82 million clinical exams performed in the U.S in 2016. Despite an average annual decline of 2% since a high of 85.3 million in 2011, it is highly sought for visualizing a host of medical conditions because of its clinical advantages in providing high spatial resolution and fast imaging time. Although limited, the high resolution of CT imaging enables small objects such as lesions to be realized with good detail. Partly due to their size and the fact that CT is noise and resolution limited, the effects of system resolution and lesion characterization processes (i.e., segmentation and CAD algorithms) are challenging to quantify. For this reason, there is a significant need to account for system resolution and algorithm impact on lesion characterization in a quantitatively reproducible manner.
Cancer is the second leading cause of death in the U.S. A fundamental aspect of cancer diagnosis, treatment and management is effective use of medical imaging. In recent years, cancer screening has received significant attention. In fact, results of screening suggest that early cancer detection can result in higher survival rates.
Beyond just visual inspection, extraction of quantitative lesion features could provide more diagnostic and treatment benefits. Assessing the quantitative capabilities of CT systems is complicated by technical factors such as noise, blur, and motion artifacts. As such, traditional modulation transfer function (MTF) methods are insufficient in characterizing system resolution, especially when non-linearities are introduced by iterative reconstruction. These aforementioned factors contribute to a major component of lesion characterization uncertainty in that they limit the apprehension of lesion ground truth. That being said, there is a wealth of quantifiable information that can be garnered from clinical images, since lesion size, morphology, and potentially texture (i.e., internal heterogeneities) are important quantitative biomarkers for effective clinical decision-making. Considering this, the imaging physics community is steadily progressing toward a quantitative paradigm in CT.
As such, the purpose of this doctoral project was to develop, validate, and disseminate a new phantom, image databases and assessment tools that are appropriate for ground truth lesion characterization in the context of modern x-ray computed tomography (CT) systems. The project developed lesion assessment methods in the framework of two distinct modes, (a) anthropomorphic phantoms and (b) clinical images.
As an alternative to the MTF, the first aspect of this project aimed at validating the task transfer function (TTF), which is a quantitative measure of system resolution. TTF was used as a means to account for the accurate modeling of low-contrast signal transfer properties of a non-linear imaging system. This study assessed the TTF as a CT system resolution model for lesion blur in the context of reconstruction algorithm, dose, and lesion size, shape, and contrast. TTF-blurred simulated lesions were compared with CT images of corresponding physical lesions using a series of comparative tools. Amidst the presence of confounding factors, in a multiple alternative forced-choice testing paradigm (4AFC) reader study, it was found that readers performed a little better than random guessing in detecting simulated lesions at a rate of 37.9±3.1% (25% implied random guessing). The visual appearance, edge-blur, size, and shape of simulated lesions were similar to the physical lesions, which suggested 3D-TTF modeled the low-contrast signal transfer properties of this non-linear CT reasonably well.
In the second study, the TTF became a useful tool for effective implementation in lesion simulation and virtual insertion. A TTF-based lesion simulation framework was developed to model lesion’s morphology in terms of size and shape. The Lungman phantom (Kyoto, Japan) was used in the implementation of two new virtual lesion insertion methods (i.e., the projection- and image-domain virtual lesion insertion methods). A third method was also used as a benchmark which was previously developed by the U.S. Food and Drug Administration (FDA). Using these TTF-based insertion methods, TTF-blurred computer aided design (CAD) lesions were virtually inserted into phantom CT projections or reconstructed data. This study compared a series of virtually-inserted, TTF-blurred CAD lesions against a corresponding series of CT-blurred physical lesions. Pair-wise comparisons were made in terms of size and shape, yielding a 3% difference in volume and a 5% difference in shape between physical and simulated lesions. This study provided indication that the proposed lesion modeling framework could quantitatively produce realistic surrogates to real lesion.
Third, a systematic assessment of bias and variability in lesion texture feature measurement was performed across a series of clinical image acquisition settings and reconstruction algorithms. A series of CT images using three computational phantoms with anatomically-informed texture were simulated representing four in-plane pixel sizes, three slice thicknesses, three dose levels, and 33 noise and resolution models, characteristic of five commercial scanners (GE LightSpeed VCT, GE Discovery 750 HD, GE Revolution, Siemens Definition Flash, and Siemens Force). 21 statistical texture features were calculated and compared between the ground truth phantom (i.e., pre-imaging) and its corresponding post-imaging simulations. Also, each texture feature was measured with four unique volumes of interest (VOIs) sizes. Across, VOI sizes and imaging settings, the percent relative difference ranged [-97%, 1230%], and the coefficient of variation ranged [1.12%, 71.79%], between the post-imaging simulation and the ground truth. The dynamic range of results indicate that image acquisition and reconstruction conditions (i.e., in-plane pixel sizes, slice thicknesses, dose levels, and reconstruction kernels) can lead to significant bias and variability in texture feature measurements. These results indicate that reconstruction and segmentation had notable effects on the bias and variability of feature measurement, thus, underscoring the need to appropriately account for system and segmentation effects on lesion characterization.
Building on the results of the TTF validation study, techniques for virtual lesion insertion study, and the texture feature assessment study, the next three studies focused on developing and validating hybrid datasets (i.e., insertion of simulated lesions into phantom and patient CT images). The fourth study was intended to determine whether interchangeability exist between real and simulated lesions in the context of patient CT images. Virtual lesions were generated based on real patient lesions extracted from the Reference Image Database to Evaluate Therapy Response (RIDER) CT dataset and were compared with their real counterparts based on lesion size. 30 pathologically-confirmed malignancies from thoracic patient CT images were modeled. Simulated lesions were re-inserted into the original CT images using the image-domain insertion program. Four readers performed volume measurements using three commercial segmentation tools. The relative volume estimation performance of segmentation tools was done to compare measures of real lesions in actual patient CT images and simulated lesions virtually inserted into the same patient images (i.e., hybrid datasets). Direct volume comparison showed consistent trends between real and simulated lesions across all segmentation algorithms, readers, and lesion shapes. Overall, there was a 5% volumetric difference between real and simulated lesions. The results of this study add to the realization of the potential applications of virtual lesions as surrogates to real clinical lesions, not just in terms of appearance, by also quantitatively.
In a fifth study, a new approach was designed to evaluate the potential for hybrid datasets with a priori known lesion volume, to serve as a replacement to clinical images in the context of segmentation algorithm compliance with the Quantitative Imaging Biomarkers Alliance (QIBA) Profile outline. This study occurred in two phases, namely a phantom and clinical phase. The phantom phase utilized the Lungman phantom and the clinical phase utilized the same base patient images from the RIDER dataset. In the phantom, hybrid datasets were generated by virtually inserting 16 simulated lesions corresponding to physical lesions into the phantom images using the projection- and image-domain (Method 1 and Method 2) techniques from the second study, along with the FDA (Method 3) technique. For the clinical data, only Method 2 was used to insert simulated lesions corresponding to real lesions. In all, across 16 participating groups, results showed that none of the virtual insertion methods were equivalent to the physical phantom based on a 5% bias margin of tolerance. However, the magnitude of this difference was small (across all groups, 2.4%, 5.4%, and 2% for Methods 1, 2, and 3, respectively).
The final aspect of this project aimed at developing hybrid datasets for use by the wider imaging community. These were composed of anthropomorphic lung and liver lesions, embedded in thoracic and abdominal images as a means to help assess lesion characterization directly from patient images. Each dataset was outfitted with a full complement of descriptive information for each inserted lesion including lesion size, shape, texture, and contrast.
In conclusion, this dissertation provides to the scientific community a new phantom, analysis techniques, modeling tools, and datasets that can aid in appropriately evaluating lesion characterization in modern CT systems. The new techniques proposed by this dissertation offer a more clinically relevant approach to assessing the impact of CT system and segmentation/CADx algorithms on lesion characterization.
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 United States License.
Rights for Collection: Duke Dissertations