Task-based transferable deep-learning scatter correction in cone beam computed tomography: a simulation study.

Purpose: X-ray scatter significantly affects the image quality of cone beam computed tomography (CBCT). Although convolutional neural networks (CNNs) have shown promise in correcting x-ray scatter, their effectiveness is hindered by two main challenges: the necessity for extensive datasets and the uncertainty regarding model generalizability. This study introduces a task-based paradigm to overcome these obstacles, enhancing the application of CNNs in scatter correction. Approach: Using a CNN with U-net architecture, the proposed methodology employs a two-stage training process for scatter correction in CBCT scans. Initially, the CNN is pre-trained on approximately 4000 image pairs from geometric phantom projections, then fine-tuned using transfer learning (TL) on 250 image pairs of anthropomorphic projections, enabling task-specific adaptations with minimal data. 2D scatter ratio (SR) maps from projection data were considered as CNN targets, and such maps were used to perform the scatter prediction. The fine-tuning process for specific imaging tasks, like head and neck imaging, involved simulating scans of an anthropomorphic phantom and pre-processing the data for CNN retraining. Results: For the pre-training stage, it was observed that SR predictions were quite accurate (SSIM≥0.9). The accuracy of SR predictions was further improved after TL, with a relatively short retraining time (≈70 times faster than pre-training) and using considerably fewer samples compared to the pre-training dataset (≈12 times smaller). Conclusions: A fast and low-cost methodology to generate task-specific CNN for scatter correction in CBCT was developed. CNN models trained with the proposed methodology were successful to correct x-ray scatter in anthropomorphic structures, unknown to the network, for simulated data.

Toward understanding deep learning classification of anatomic sites: lessons from the development of a CBCT projection classifier.

Purpose: Deep learning (DL) applications strongly depend on the training dataset and convolutional neural network architecture; however, it is unclear how to objectively select such parameters. We investigate the classification performance of different DL models and training schemes for the anatomic classification of cone-beam computed tomography (CBCT) projections. Approach: CBCT scans from 1055 patients were collected and manually classified into five anatomic classes and used to develop DL models to predict the anatomic class from single x-ray projections. VGG-16, Xception, and Inception v3 architectures were trained with 75% of the data, and the remaining 25% was used for testing and evaluation. To study the dependence of the classification performance on dataset size, training data was downsampled to various dataset sizes. Gradient-weighted class activation maps (grad-CAM) were generated using the model with highest classification performance, to identify regions with strong influence on CNN decisions. Results: The highest precision and recall values were achieved with VGG-16. One of the best performing combinations was the VGG-16 trained with 90 deg projections (mean class precision = 0.87). The training dataset size could be reduced to ∼ 50 % of its initial size, without compromising the classification performance. For correctly classified cases, Grad-CAM were more heavily weighted for anatomically relevant regions. Conclusions: It was possible to determine those dependencies with a higher influence on the classification performance of DL models for the studied task. Grad-CAM enabled the identification of possible sources of class confusion.

Development of a neonate X-ray phantom for 2D imaging applications using single-tone inkjet printing.

PURPOSE: Inkjet printers can be used to fabricate anthropomorphic phantoms by the use of iodine-doped ink. However, challenges persist in implementing this technique. The calibration from grayscale to ink density is complex and time-consuming. The purpose of this work is to develop a printing methodology that requires a simpler calibration and is less dependent on printer characteristics to produce the desired range of x-ray attenuation values. METHODS: Conventional grayscale printing was substituted by single-tone printing; that is, the superposition of pure black layers of iodinated ink. Printing was performed with a consumer-grade inkjet printer using ink made of potassium-iodide (KI) dissolved in water at 1 g/ml. A calibration for the attenuation of ink was measured using a commercial x-ray system at 70 kVp. A neonate radiograph obtained at 70 kVp served as an anatomical model. The attenuation map of the neonate radiograph was processed into a series of single-tone images. Single-tone images were printed, stacked, and imaged at 70 kVp. The phantom was evaluated by comparing attenuation values between the printed phantom and the original radiograph; attenuation maps were compared using the structural similarity index measure (SSIM), while attenuation histograms were compared using the Kullback-Leibler (KL) divergence. A region of interest (ROI)-based analysis was also performed, where the attenuation distribution within given ROIs was compared between phantom and patient. The phantom sharpness was evaluated in terms of modulation transfer function (MTF) estimates and signal spread profiles of high spatial resolution features in the image. RESULTS: The printed phantom required 36 pages. The printing queue was automated and it took about 2 h to print the phantom. The radiograph of the printed phantom demonstrated a close resemblance to the original neonate radiograph. The SSIM of the phantom with respect to that of the patient was 0.53. Both patient and phantom attenuation histograms followed similar distributions, and the KL divergence between such histograms was 0.20. The ROI-based analysis showed that the largest deviations from patient attenuation values were observed at the higher and lower ends of the attenuation range. The limiting resolution of the proposed methodology was about 1 mm. CONCLUSION: A methodology to generate a neonate phantom for 2D imaging applications, using single-tone printing, was developed. This method only requires a single-value calibration and required less than 2 h to print a complete phantom.

A signal detection model for quantifying overregularization in nonlinear image reconstruction.

Many useful image quality metrics for evaluating linear image reconstruction techniques do not apply to or are difficult to interpret for nonlinear image reconstruction. The vast majority of metrics employed for evaluating nonlinear image reconstruction are based on some form of global image fidelity, such as image root mean square error (RMSE). Use of such metrics can lead to overregularization in the sense that they can favor removal of subtle details in the image. To address this shortcoming, we develop an image quality metric based on signal detection that serves as a surrogate to the qualitative loss of fine image details. The metric is demonstrated in the context of a breast CT simulation, where different equal-dose configurations are considered. The configurations differ in the number of projections acquired. Image reconstruction is performed with a nonlinear algorithm based on total variation constrained least-squares (TV-LSQ). The resulting images are studied as a function of three parameters: number of views acquired, total variation constraint value, and number of iterations. The images are evaluated visually, with image RMSE, and with the proposed signal-detection-based metric. The latter uses a small signal, and computes detectability in the sinogram and in the reconstructed image. Loss of signal detectability through the image reconstruction process is taken as a quantitative measure of loss of fine details in the image. Loss of signal detectability is seen to correlate well with the blocky or patchy appearance due to overregularization with TV-LSQ, and this trend runs counter to the image RMSE metric, which tends to favor the over-regularized images. The proposed signal detection-based metric provides an image quality assessment that is complimentary to that of image RMSE. Using the two metrics in concert may yield a useful prescription for determining CT algorithm and configuration parameters when nonlinear image reconstruction is used.

Impact of noise reduction schemes on quantitative accuracy of CT numbers.

PURPOSE: In previous works, it has been demonstrated that for filtered backprojection (FBP) reconstruction-based computed tomography (CT) images, the measured CT numbers are biased and the bias level decreases with increasing radiation dose. Low-dose scans typically include noise reduction schemes to reduce noise level. The purpose of this work was to investigate the potential impact of different noise reduction schemes on the CT number bias. METHODS: Three different filtration methods: Gaussian, adaptive trimmed mean (ATM), and anisotropic diffusion (AD) were implemented to reduce noise. All filters were independently applied in three different domains: raw counts, log-processed sinogram, or reconstructed image domain. A quality assurance phantom was scanned on a benchtop CT cone beam CT system, at dose levels ranging from 0.6 to 4.0 mGy. The conventional FBP reconstructions were performed to reconstruct CT images for the study of CT number biases. The CT number bias of different material inserts in the phantom was then measured. To further study the overall impact of CT number bias together with the potential consequences of noise reduction schemes on both the spatial resolution and noise characteristics, the task-based detectability of a high-contrast and high spatial resolution imaging task was used as an example to assess the performance of each noise reduction scheme. To qualitatively assess the impact of these noise reduction schemes on image, an anthropomorphic head phantom was also scanned on the benchtop CT system and processed with the above noise reduction schemes to generate images for demonstration. RESULTS: Our results demonstrated the following major findings: (a) CT number bias can be significantly reduced when the noise reduction schemes are implemented in the raw counts domain; CT number bias cannot be reduced when these noise reduction schemes are implemented either in the reconstructed image domain or in the log-processed sinogram domain. (b) The extent of CT number bias reduction is dependent on both the material composition and noise reduction parameters. (c) The overall impact of the noise reduction schemes can be studied using the task-based detectability analysis framework and this framework can be used to select the appropriate parameters in each noise reduction scheme to optimize the performance for a given imaging task. CONCLUSIONS: Noise reduction schemes can be used to considerably reduce CT number bias when they are implemented in the raw counts domain; however, their application cannot be arbitrarily extended to either the log-processed sinogram data domain or image domain. Trade-offs between bias reduction and overall image quality must be studied for an optimal performance of a given imaging task.

Quantitative accuracy of CT numbers: Theoretical analyses and experimental studies.

PURPOSE: The CT number accuracy, that is, CT number bias, plays an important role in clinical diagnosis. When strategies to reduce radiation dose are discussed, it is important to make sure that the CT number bias is controlled within an acceptable range. The purpose of this paper was to investigate the dependence of CT number bias on radiation dose level and on image contrast (i.e., the difference in CT number between the ROI and the background) in Computed Tomography (CT). METHODS: A lesion-background model was introduced to theoretically study how the CT number bias changes with radiation exposure level and with CT number contrast when a simple linear reconstruction algorithm such as filtered backprojection (FBP) is used. The theoretical results were validated with experimental studies using a benchtop CT system equipped with a photon-counting detector (XC-HYDRA FX50, XCounter AB, Sweden) and a clinical diagnostic MDCT scanner (Discovery CT750 HD, GE Healthcare, Waukesha, WI, USA) equipped with an energy-integrating detector. The Catphan phantom (Catphan 600, the Phantom Laboratory, Salem, NY, USA) was scanned at different mAs levels and 50 scans were performed for each mAs. The bias of CT number was evaluated for each combination of mAs and ROIs with different contrast levels. An anthropomorphic phantom (ATOM 10-year-old phantom, Model 706, CIRS Inc. Norfolk, VA, USA) with much more heterogeneous object content was used to test the applicability of the theory to the more general image object cases. RESULTS: Both theoretical and experimental studies showed that the CT number bias is inversely proportional to the radiation exposure level yet linearly dependent on the CT number contrast between the lesion and the background, that is, Bias ( µ ^ 1 FBP ) = α mAs ( 1 + ß Δ H U ) . CONCLUSIONS: The quantitative accuracy of CT numbers can be problematic and thus needs some extra attention when radiation dose is reduced. In this work, we showed that the bias of the FBP reconstruction increases as mAs is reduced; both positive and negative bias can be observed depending on the contrast difference between a targeted ROI and its surrounding background tissues.

Modified ideal observer model (MIOM) for high-contrast and high-spatial resolution CT imaging tasks.

PURPOSE: Although a variety of mathematical observer models have been developed to predict human observer performance for low contrast lesion detection tasks, their predictive power for high contrast and high spatial resolution discrimination imaging tasks, including those in CT bone imaging, could be limited. The purpose of this work was to develop a modified observer model that has improved correlation with human observer performance for these tasks. METHODS: The proposed observer model, referred to as the modified ideal observer model (MIOM), uses a weight function to penalize components in the task function that have less contribution to the actual human observer performance for high contrast and high spatial resolution discrimination tasks. To validate MIOM, both human observer and observer model studies were performed, each using exactly the same CT imaging task [discrimination of a connected component in a high contrast (1000 HU) high spatial resolution bone fracture model (0.3 mm)] and experimental CT image data. For the human observer studies, three physicist observers rated the connectivity of the fracture model using a five-point Likert scale; for the observer model studies, a total of five observer models, including both conventional models and the proposed MIOM, were used to calculate the discrimination capability of the CT images in resolving the connected component. Images used in the studies encompassed nine different reconstruction kernels. Correlation between human and observer model performance for these kernels were quantified using the Spearman rank correlation coefficient (ρ). After the validation study, an example application of MIOM was presented, in which the observer model was used to select the optimal reconstruction kernel for a High-Resolution (Hi-Res, GE Healthcare) CT scan technique. RESULTS: The performance of the proposed MIOM correlated well with that of the human observers with a Spearman rank correlation coefficient ρ of 0.88 (P = 0.003). In comparison, the value of ρ was 0.05 (P = 0.904) for the ideal observer, 0.05 (P = 0.904) for the non-prewhitening observer, -0.18 (P = 0.634) for the non-prewhitening observer with eye filter and internal noise, and 0.30 (P = 0.427) for the prewhitening observer with eye filter and internal noise. Using the validated MIOM, the optimal reconstruction kernel for the Hi-Res mode to perform high spatial resolution and high contrast discrimination imaging tasks was determined to be the HD Ultra kernel at the center of the scan field of view (SFOV), or the Lung kernel at the peripheral region of the SFOV. This result was consistent with visual observations of nasal CT images of an in vivo canine subject. CONCLUSION: Compared with other observer models, the proposed modified ideal observer model provides significantly improved correlation with human observers for high contrast and high spatial resolution CT imaging tasks.

Clinical study of contrast-enhanced digital mammography and the evaluation of blood and lymphatic microvessel density.

OBJECTIVE: To correlate image parameters in contrast-enhanced digital mammography (CEDM) with blood and lymphatic microvessel density (MVD). METHODS: 18 Breast Imaging-Reporting and Data System (BI-RADS)-4 to BI-RADS-5 patients were subjected to CEDM. Craniocaudal views were acquired, two views (low and high energy) before iodine contrast medium (CM) injection and four views (high energy) 1-5 min afterwards. Processing included registration and two subtraction modalities, traditional single-energy temporal (high-energy) and "dual-energy temporal with a matrix", proposed to improve lesion conspicuity. Images were calibrated into iodine thickness, and iodine uptake, contrast, time-intensity and time-contrast kinetic curves were quantified. Image indicators were compared with MVD evaluated by anti-CD105 and anti-podoplanin (D2-40) immunohistochemistry. RESULTS: 11 lesions were cancerous and 7 were benign. CEDM subtraction strongly increased conspicuity of lesions enhanced by iodine uptake. A strong correlation was observed between lymphatic vessels and blood vessels; all benign lesions had <30 blood microvessels per field, and all cancers had more than this value. MVD showed no correlation with iodine uptake, nor with contrast. The most frequent curve was early uptake followed by plateau for uptake and contrast in benign and malignant lesions. The positive-predictive value of uptake dynamics was 73% and that of contrast was 64%. CONCLUSION: CEDM increased lesion visibility and showed additional features compared with conventional mammography. Lack of correlation between image parameters and MVD is probably due to tumour tissue heterogeneity, mammography projective nature and/or dependence of extracellular iodine irrigation on tissue composition. ADVANCES IN KNOWLEDGE: Quantitative analysis of CEDM images was performed. Image parameters and MVD showed no correlation. Probably, this is indication of the complex dependence of CM perfusion on tumour microenvironment.

Hi-Res scan mode in clinical MDCT systems: Experimental assessment of spatial resolution performance.

PURPOSE: The introduction of a High-Resolution (Hi-Res) scan mode and another associated option that combines Hi-Res mode with the so-called High Definition (HD) reconstruction kernels (referred to as a Hi-Res/HD mode in this paper) in some multi-detector CT (MDCT) systems offers new opportunities to increase spatial resolution for some clinical applications that demand high spatial resolution. The purpose of this work was to quantify the in-plane spatial resolution along both the radial direction and tangential direction for the Hi-Res and Hi-Res/HD scan modes at different off-center positions. METHODS: A technique was introduced and validated to address the signal saturation problem encountered in the attempt to quantify spatial resolution for the Hi-Res and Hi-Res/HD scan modes. Using the proposed method, the modulation transfer functions (MTFs) of a 64-slice MDCT system (Discovery CT750 HD, GE Healthcare) equipped with both Hi-Res and Hi-Res/HD modes were measured using a metal bead at nine different off-centered positions (0-16 cm with a step size of 2 cm); at each position, both conventional scans and Hi-Res scans were performed. For each type of scan and position, 80 repeated acquisitions were performed to reduce noise induced uncertainties in the MTF measurements. A total of 15 reconstruction kernels, including eight conventional kernels and seven HD kernels, were used to reconstruct CT images of the bead. An ex vivo animal study consisting of a bone fracture model was performed to corroborate the MTF results, as the detection of this high-contrast and high frequency task is predominantly determined by spatial resolution. Images of this animal model generated by different scan modes and reconstruction kernels were qualitatively compared with the MTF results. RESULTS: At the centered position, the use of Hi-Res mode resulted in a slight improvement in the MTF; each HD kernel generated higher spatial resolution than its counterpart conventional kernel. However, the MTF along the tangential direction of the scan field of view (SFOV) was significantly degraded at off-centered positions, yet the combined Hi-Res/HD mode reduced this azimuthal MTF degradation. Images of the animal bone fracture model confirmed the improved spatial resolution at the off-centered positions through the use of the Hi-Res mode and HD kernels. CONCLUSIONS: The Hi-Res/HD scan improve spatial resolution of MDCT systems at both centered and off-centered positions.