Adrian A. Sanchez

Comparison of human and Hotelling observer performance for a fan-beam CT signal detection task.

PURPOSE A human observer study was performed for a signal detection task for the case of fan-beam x-ray computed tomography. Hotelling observer (HO) performance was calculated for the same detection task without the use of efficient channels. By considering the full image covariance produced by the filtered backprojection (FBP) algorithm and avoiding the use of channels in the computation of HO performance, the authors establish an absolute upper bound on signal detectability. Therefore, this study serves as a baseline for relating human and ideal observer performance in the case of fan-beam CT. METHODS Eight human observers participated in a two-alternative forced choice experiment where the signal of interest was a small simulated ellipsoid in the presence of independent, identically distributed Gaussian detector noise. Theoretical performance of the HO, which is equivalent to the ideal observer in this case (see Sec. 13.2.12 in Barrett and Myers [Foundations of Image Science (Wiley, Hoboken, NJ, 2004)], was also computed and compared to the performance of the human observers. In addition to a reference FBP implementation, two FBP implementations with inherent loss of HO signal detectability (e.g., by apodizing the ramp filter) were also investigated. Each of these latter two implementations takes the form of a discrete-to-discrete linear operator (i.e., a matrix), which has a nontrivial null-space resulting in the loss of detectability. RESULTS Estimated observer detectability index (d(A)) values for the human observers and SNR values for the HO were obtained. While Hanning filtering in the FBP implementation with a cutoff frequency of 1/4 of the Nyquist frequency reduces HO SNR (due to the reconstruction matrixs nontrivial null-space), this filtering was shown to consistently improve human observer performance. By contrast, increasing the image pixel size was seen to have a comparable effect on both the HO and the human observers, degrading performance. CONCLUSIONS These results, which characterize HO and human observer performance for a signal detection task in fan-beam FBP noise, form a basis for applying model observer metrics to fan-beam CT when knowledge of the full image-domain noise statistics is important. Further, by calculating HO performance without relying on channels, these results are particularly relevant when an information theoretic approach is considered, e.g., in optimization of the image reconstruction algorithm with respect to preservation of signal detectability. Finally, the HO (which is here equivalent to the ideal observer) provides an absolute upper bound on detection performance, and our results therefore provide insight into the performance of human observers relative to the optimum for two different cases wherein ideal observer performance is compromised through degradation of the data quality. In one case (regularization), human performance is improved to practically ideal performance, and in the other (larger pixel size), ideal and human observer performance are approximately degraded equivalently.

Task-based optimization of dedicated breast CT via Hotelling observer metrics

PURPOSE The purpose of this work is to develop and demonstrate a set of practical metrics for CT systems optimization. These metrics, based on the Hotelling observer (HO) figure of merit, are task-based. The authors therefore take the specific example of optimizing a dedicated breast CT system, including the reconstruction algorithm, for two relevant tasks, signal detection and Rayleigh discrimination. METHODS A dedicated breast CT system is simulated using specifications in the literature from an existing prototype. The authors optimize configuration and image reconstruction algorithm parameters for two tasks: the detection of simulated microcalcifications and the discrimination of two adjacent, high-contrast signals, known as the Rayleigh discrimination task. The effects on task performance of breast diameter, signal location, image grid size, projection view number, and reconstruction filter were all investigated. Two HO metrics were evaluated: the percentage of correct decisions in a two-alternative forced choice experiment (equivalent to area under the ROC curve or AUC), and the HO efficiency, defined as the squared ratio of HO signal-to-noise ratio (SNR) in the reconstructed image to HO SNR in the projection data. RESULTS The ease and efficiency of the HO metric computation allows a rapid high-resolution survey of many system parameters. Optimization of a range of system parameters using the HO results in images that subjectively appear optimal for the tasks investigated. Further, the results of assessment through the HO reproduce closely many existing results in the literature regarding the impact of parameter selection on image quality. CONCLUSIONS This study demonstrates the utility of a task-based approach to system design, evaluation, and optimization. The methodology presented is equally applicable to determining the impact of a wide range of factors, including patient parameters, system and acquisition design, and the reconstruction algorithm. The results demonstrate the versatility of the proposed HO formalism by not only generating a set of parameters that are optimal for a given task but also by qualitatively reproducing many existing results from the breast CT literature. Meanwhile, the implementation of the proposed methodology is straightforward and entirely simulation-based. This is an attractive feature for many system optimization problems, where the goal is to analyze the individual system components such as the image reconstruction algorithm. Final assessment of the system as a whole should be based also on real data studies.

Investigating simulation‐based metrics for characterizing linear iterative reconstruction in digital breast tomosynthesis

Purpose Simulation‐based image quality metrics are adapted and investigated for characterizing the parameter dependences of linear iterative image reconstruction for DBT. Methods Three metrics based on a 2D DBT simulation are investigated: (1) a root‐mean‐square‐error (RMSE) between the test phantom and reconstructed image, (2) a gradient RMSE where the comparison is made after taking a spatial gradient of both image and phantom, and (3) a region‐of‐interest (ROI) Hotelling observer (HO) for signal‐known‐exactly/background‐known‐exactly (SKE/BKE) and signal‐known‐exactly/background‐known‐statistically (SKE/BKS) detection tasks. Two simulation studies are performed using the aforementioned metrics, varying voxel aspect ratio, and regularization strength for two types of Tikhonov‐regularized least‐squares optimization. The RMSE metrics are applied to a 2D test phantom with resolution bar patterns at varying angles, and the ROI‐HO metric is applied to two tasks relevant to DBT: lesion detection, modeled by use of a large, low‐contrast signal, and microcalcification detection, modeled by use of a small, high‐contrast signal. The RMSE metric trends are compared with visual assessment of the reconstructed bar‐pattern phantom. The ROI‐HO metric trends are compared with 3D reconstructed images from ACR phantom data acquired with a Hologic Selenia Dimensions DBT system. Results Sensitivity of the image RMSE to mean pixel value is found to limit its applicability to the assessment of DBT image reconstruction. The image gradient RMSE is insensitive to mean pixel value and appears to track better with subjective visualization of the reconstructed bar‐pattern phantom. The ROI‐HO metric shows an increasing trend with regularization strength for both forms of Tikhonov‐regularized least‐squares; however, this metric saturates at intermediate regularization strength indicating a point of diminishing returns for signal detection. Visualization with the reconstructed ACR phantom images appear to show a similar dependence with regularization strength. Conclusions From the limited studies presented it appears that image gradient RMSE trends correspond with visual assessment better than image RMSE for DBT image reconstruction. The ROI‐HO metric for both detection tasks also appears to reflect visual trends in the ACR phantom reconstructions as a function of regularization strength. We point out, however, that the true utility of these metrics can only be assessed after amassing more data.

Use of the Hotelling observer to optimize image reconstruction in digital breast tomosynthesis

Abstract. We propose an implementation of the Hotelling observer that can be applied to the optimization of linear image reconstruction algorithms in digital breast tomosynthesis. The method is based on considering information within a specific region of interest, and it is applied to the optimization of algorithms for detectability of microcalcifications. Several linear algorithms are considered: simple back-projection, filtered back-projection, back-projection filtration, and Λ-tomography. The optimized algorithms are then evaluated through the reconstruction of phantom data. The method appears robust across algorithms and parameters and leads to the generation of algorithm implementations which subjectively appear optimized for the task of interest.

Region of interest based Hotelling observer for computed tomography with comparison to alternative methods

Abstract. We compare several approaches to estimation of Hotelling observer (HO) performance in x-ray computed tomography (CT). We consider the case where the signal of interest is small so that the reconstructed image can be restricted to a small region of interest (ROI) surrounding the signal. This reduces the dimensionality of the image covariance matrix so that direct computation of HO metrics within the ROI is feasible. We propose that this approach is directly applicable to systems optimization in CT; however, many alternative approaches exist, which make computation of HO performance tractable through a range of approximations, assumptions, or estimation strategies. Here, we compare several of these methods, including the use of Laguerre-Gauss channels, discrete Fourier domain computation of the HO (which assumes noise stationarity), and two approaches to HO estimation through samples of noisy images. Since our method computes HO performance exactly within an ROI, this allows us to investigate the validity of the assumptions inherent in various common approaches to HO estimation, such as the stationarity assumption in the case of the discrete Fourier transform domain method.

Efficient iterative image reconstruction algorithm for dedicated breast CT

Dedicated breast computed tomography (bCT) is currently being studied as a potential screening method for breast cancer. The X-ray exposure is set low to achieve an average glandular dose comparable to that of mammography, yielding projection data that contains high levels of noise. Iterative image reconstruction (IIR) algorithms may be well-suited for the system since they potentially reduce the effects of noise in the reconstructed images. However, IIR outcomes can be difficult to control since the algorithm parameters do not directly correspond to the image properties. Also, IIR algorithms are computationally demanding and have optimal parameter settings that depend on the size and shape of the breast and positioning of the patient. In this work, we design an efficient IIR algorithm with meaningful parameter specifications and that can be used on a large, diverse sample of bCT cases. The flexibility and efficiency of this method comes from having the final image produced by a linear combination of two separately reconstructed images - one containing gray level information and the other with enhanced high frequency components. Both of the images result from few iterations of separate IIR algorithms. The proposed algorithm depends on two parameters both of which have a well-defined impact on image quality. The algorithm is applied to numerous bCT cases from a dedicated bCT prototype system developed at University of California, Davis.

Evaluation of spectral CT data acquisition methods via non-stochastic variance maps

Recently, photon counting detectors capable of extracting spectral information have received much attention in CT, with promise of using spectral information to construct material basis images, correct beam-hardening artifacts, or provide improved imaging of K-edge contrast agents.1, 2 In this work, we focus on the goal of constructing images of basis material maps, and investigate the feasibility of analytically computing pixel variance maps for these images, so that alternative data acquisition and reconstruction methods can be compared and evaluated with respect to their noise properties. Our approach is based on linearization of the basis material decomposition and reconstruction operations, and we therefore demonstrate the method using the ubiquitous filtered back-projection algorithm, which is linear. We then performed preliminary investigation of the method by comparing basis material variance maps for two data acquisition methods that were previously found to have different noise properties:3 two-sided bin measurements acquired from separate, independent data realizations and two-sided bin measurements acquired from a single data realization.

Task-based optimization of image reconstruction in breast CT

We demonstrate a task-based assessment of image quality in dedicated breast CT in order to optimize the number of projection views acquired. The methodology we employ is based on the Hotelling Observer (HO) and its associated metrics. We consider two tasks: the Rayleigh task of discerning between two resolvable objects and a single larger object, and the signal detection task of classifying an image as belonging to either a signalpresent or signal-absent hypothesis. HO SNR values are computed for 50, 100, 200, 500, and 1000 projection view images, with the total imaging radiation dose held constant. We use the conventional fan-beam FBP algorithm and investigate the effect of varying the width of a Hanning window used in the reconstruction, since this affects both the noise properties of the image and the under-sampling artifacts which can arise in the case of sparse-view acquisitions. Our results demonstrate that fewer projection views should be used in order to increase HO performance, which in this case constitutes an upper-bound on human observer performance. However, the impact on HO SNR of using fewer projection views, each with a higher dose, is not as significant as the impact of employing regularization in the FBP reconstruction through a Hanning filter.

Utilizing the Hotelling template as a tool for CT image reconstruction algorithm design

Design of image reconstruction algorithms for CT can be significantly aided by useful metrics of image quality. Useful metrics, however, are difficult to develop due to the high-dimensionality of the CT imaging system, lack of spatial invariance in the imaging system, and a high degree of correlation among the image voxels. Although true task-based evaluation on realistic imaging tasks can be time-consuming, and a given task may be insensitive to the image reconstruction algorithm, task-based metrics can still prove useful in many contexts. For example, model observers that mimic performance of the imaging system on specific tasks can provide a low-dimensional measure of image quality while still accounting for many of the salient properties of the system and object being scanned. In this work, ideal observer performance is computed on a single detection task. The modeled signal for detection is taken to be very small - size on the order of a detector bin - and inspection of the accompanying Hotelling template is suggested. We hypothesize that improved detection on small signals may be sensitive to the reconstruction algorithm. Further, we hypothesize that structurally simple Hotelling templates may correlate with high human observer performance.

TU-FG-209-11: Validation of a Channelized Hotelling Observer to Optimize Chest Radiography Image Processing for Nodule Detection: A Human Observer Study

PURPOSE To validate the use of a Channelized Hotelling Observer (CHO) model for guiding image processing parameter selection and enable improved nodule detection in digital chest radiography. METHODS In a previous study, an anthropomorphic chest phantom was imaged with and without PMMA simulated nodules using a GE Discovery XR656 digital radiography system. The impact of image processing parameters was then explored using a CHO with 10 Laguerre-Gauss channels. In this work, we validate the CHOs trend in nodule detectability as a function of two processing parameters by conducting a signal-known-exactly, multi-reader-multi-case (MRMC) ROC observer study. Five naive readers scored confidence of nodule visualization in 384 images with 50% nodule prevalence. The image backgrounds were regions-of-interest extracted from 6 normal patient scans, and the digitally inserted simulated nodules were obtained from phantom data in previous work. Each patient image was processed with both a near-optimal and a worst-case parameter combination, as determined by the CHO for nodule detection. The same 192 ROIs were used for each image processing method, with 32 randomly selected lung ROIs per patient image. Finally, the MRMC data was analyzed using the freely available iMRMC software of Gallas et al. RESULTS The image processing parameters which were optimized for the CHO led to a statistically significant improvement (p=0.049) in human observer AUC from 0.78 to 0.86, relative to the image processing implementation which produced the lowest CHO performance. CONCLUSION Differences in user-selectable image processing methods on a commercially available digital radiography system were shown to have a marked impact on performance of human observers in the task of lung nodule detection. Further, the effect of processing on humans was similar to the effect on CHO performance. Future work will expand this study to include a wider range of detection/classification tasks and more observers, including experienced chest radiologists.