Search

Background: Deep artificial neural networks such as convolutional neural networks (CNNs) have been shown to be effective models for reducing noise in CT images while preserving anatomic details.A practical bottleneck for developing CNN-based denoising models is the procurement of training data consisting of paired examples of high-noise and low-noise CT images. Obtaining these paired data are not practical in a clinical setting where the raw projection data is not available. This work outlines a technique to optimize CNN denoising models using methods that are available in a routine clinical setting. Purpose: To demonstrate a phantom-based training framework for CNN noise reduction that can be efficiently implemented on any CT scanner. Methods: The phantom-based training framework uses supervised learning in which training data are synthesized using an image-based noise insertion technique. Ten patient image series were used for training and validation (9:1) and noise-only images obtained from anthropomorphic phantom scans. Phantom noise-only images were superimposed on patient images to imitate low-dose CT images for use in training. A modified U-Net architecture was used with mean-squared-error and feature reconstruction loss. The training framework was tested for clinically indicated whole-body-low-dose CT images, as well as routine abdomen-pelvis exams for which projection data was unavailable. Performance was assessed based on root-mean-square error, structural similarity, line profiles, and visual assessment. Results: When the CNN was tested on five reserved quarter-dose whole-body-low-dose CT images, noise was reduced by 75%, root-mean-square-error reduced by 34%, and structural similarity increased by 60%. Visual analysis and line profiles indicated that the method significantly reduced noise while maintaining spatial resolution of anatomic features. Conclusion: The proposed phantom-based training framework demonstrated strong noise reduction while preserving spatial detail. Because this method is based within the image domain, it can be easily implemented without access to projection data. [ABSTRACT FROM AUTHOR]

SynthesiZed Improved Resolution and Concurrent nOise reductioN (ZIRCON) is a multi-kernel synthesis method that creates a single series of thin-slice computed tomography (CT) images displaying low noise and high spatial resolution, increasing reader efficiency and minimizing partial volume averaging.To compare the diagnostic performance of a single set of ZIRCON images to two routine clinical image series using conventional CT head and bone reconstruction kernels for diagnosing intracranial findings and fractures in patients with trauma or suspected acute neurologic deficit.In total, 50 patients underwent clinically indicated head CT in the ER (15 normal, 35 abnormal cases). A non-reader neuroradiologist established the reference standard. Three neuroradiologists reviewed two routine clinical series (head and bone kernels) and a single ZIRCON series, detecting intracranial findings or fractures and rating confidence (0–100). Sensitivity, specificity, and jackknife free-response receiver operating characteristic (JAFROC) figure of merit (FOM) were compared (limit of non-inferiority: −0.10).ZIRCON and conventional images demonstrated comparable performance for fractures (sensitivity: 51.5% vs. 54.5%; specificity: 40.2% vs. 34.2%) and intracranial findings (sensitivity: 88.2% vs. 91.4%; specificity: 77.2% vs. 73.7%).The estimated difference of JAFROC FOM demonstrated ZIRCON non-inferiority for acute pathologies overall (0.003 [95% CI=−0.051–0.057]) and fractures (0.048 [95% CI=−0.050–0.145]) but not for intracranial findings alone (−0.024 [95% CI=−0.100–0.052]).Thin-slice, low noise, and high spatial resolution images can be created to display intracranial findings and fractures replacing multiple images series in head CT with similar performance. Future studies in more patients and further algorithmic development are warranted. [ABSTRACT FROM AUTHOR]

Background and Purpose: Photon-counting detector CT (PCD-CT) is now clinically available and offers ultra-high-resolution (UHR) imaging. Our purpose was to prospectively evaluate the relative image quality and impact on diagnostic confidence of head CTA images acquired by using a PCD-CT compared with an energy-integrating detector CT (EID-CT)., Materials and Methods: Adult patients undergoing head CTA on EID-CT also underwent a PCD-CT research examination. For both CT examinations, images were reconstructed at 0.6 mm by using a matched standard resolution (SR) kernel. Additionally, PCD-CT images were reconstructed at the thinnest section thickness of 0.2 mm (UHR) with the sharpest kernel, and denoised with a deep convolutional neural network (CNN) algorithm (PCD-UHR-CNN). Two readers (R1, R2) independently evaluated image quality in randomized, blinded fashion in 2 sessions, PCD-SR versus EID-SR and PCD-UHR-CNN versus EID-SR. The readers rated overall image quality (1 [worst] to 5 [best]) and provided a Likert comparison score (-2 [significantly inferior] to 2 [significantly superior]) for the 2 series when compared side-by-side for several image quality features, including visualization of specific arterial segments. Diagnostic confidence (0-100) was rated for PCD versus EID for specific arterial findings, if present., Results: Twenty-eight adult patients were enrolled. The volume CT dose index was similar (EID: 37.1 ± 4.7 mGy; PCD: 36.1 ± 4.0 mGy). Overall image quality for PCD-SR and PCD-UHR-CNN was higher than EID-SR (eg, PCD-UHR-CNN versus EID-SR: 4.0 ± 0.0 versus 3.0 ± 0.0 (R1), 4.9 ± 0.3 versus 3.0 ± 0.0 (R2); all P values < .001). For depiction of arterial segments, PCD-SR was preferred over EID-SR (R1: 1.0-1.3; R2: 1.0-1.8), and PCD-UHR-CNN over EID-SR (R1: 0.9-1.4; R2: 1.9-2.0). Diagnostic confidence of arterial findings for PCD-SR and PCD-UHR-CNN was significantly higher than EID-SR: eg, PCD-UHR-CNN versus EID-SR: 93.0 ± 5.8 versus 78.2 ± 9.3 (R1), 88.6 ± 5.9 versus 70.4 ± 5.0 (R2); all P values < .001., Conclusions: PCD-CT provides improved image quality for head CTA images compared with EID-CT, both when PCD and EID reconstructions are matched, and to an even greater extent when PCD-UHR reconstruction is combined with a CNN denoising algorithm., (© 2024 by American Journal of Neuroradiology.)

The Rose criterion, stating that an object is detectable if it is five standard deviations above background, has been used as a rule of thumb for decades but its applicability is limited in computed tomography. Recent denoising algorithms, powered by convolutional neural networks, promise to reveal objects that were previously obscured by noise, but any denoising algorithm is fundamentally limited by the statistics of the sinogram. In this work, we estimate the minimum SNR necessary for detecting one of a set of objects in the projection domain. We assume there is a set of objects O for which detection is desired, and we study an ideal observer that sequentially compares each member of O to the null hypothesis. This comparison can be reduced to the classic one-dimensional signal detection problem between two Gaussians with different mean values, and from this we define a quantity, the projection SNR. We use simulations to estimate the minimum projection SNR necessary to achieve a sensitivity of 80% and specificity of 80%. We find that when we model a search task of a circular 6 mm lesion in a region of interest that is 60 mm by 60 mm by 10 slices, the minimum projection SNR is 5.1. This required SNR is reminiscent of the Rose criterion but is derived with entirely different assumptions, including the application of the ideal observer in the projection domain.

In this study, we describe a systematic approach to optimize deep-learning-based image processing algorithms using random search. The optimization technique is demonstrated on a phantom-based noise reduction training framework; however, the techniques described can be applied generally for other deep learning image processing applications. The parameter space explored included number of convolutional layers, number of filters, kernel size, loss function, and network architecture (either U-Net or ResNet). A total of 100 network models were examined (50 random search, 50 ablation experiments). Following the random search, ablation experiments resulted in a very minor performance improvement indicating near optimal settings were found during the random search. The top performing network architecture was a U-Net with 4 pooling layers, 64 filters, 3×3 kernel size, ELU activation, and a weighted feature reconstruction loss (0.2×VGG + 0.8×MSE). Relative to the low-dose input image, the CNN reduced noise by 90%, reduced RMSE by 34%, and increased SSIM by 76% on six patient exams reserved for testing. The visualization of hepatic and bone lesions was greatly improved following noise reduction.