Performance of a deep learning-based CT image denoising method: Generalizability over dose, reconstruction kernel and slice thickness

Utility of block-matching and 3D filter for reproducibility of lung density and denoising in low-dose chest CT: A pilot study

PURPOSE

This study aimed to acquire an image quality consistent with that of full-dose chest computed tomography (CT) when obtaining low-dose chest CT images and to analyze the effects of block-matching and 3D (BM3D) filters on lung density measurements and noise reduction in lung parenchyma.

METHODS

Using full-dose chest CT images, we evaluated lung density measurements and noise reduction in lung parenchyma images for low-dose chest CT. Three filters (median, Wiener, and the proposed BM3D) were applied to low-dose chest CT images for comparison and analysis with images from full-dose chest CT. To evaluate lung density measurements, we measured CT attenuation at the 15th percentile of the lung CT histogram. The coefficient of variation (COV) and contrast-to-noise ratio (CNR) were used to evaluate the noise level.

RESULTS

The 15th percentile of the lung CT histogram showed the smallest difference between full- and low-dose CT when applying the BM3D filter, and the highest difference between full- and low-dose CT without filters (full-dose =  - 926.28 ± 0.32, BM3D =  - 926.65 ± 0.32, and low-dose =  - 959.43 ± 0.95) (p < 0.05). The COV was smallest when applying the BM3D filter, whereas the CNR was the highest (p < 0.05).

CONCLUSIONS

The results of the study prove that the BM3D filter can reduce image noise while increasing the reproducibility of the lung density, even for low-dose chest CT.

Multiple Kernel Synthesis of Head CT Using a Task-Based Loss Function

In CT imaging of the head, multiple image series are routinely reconstructed with different kernels and slice thicknesses. Reviewing the redundant information is an inefficient process for radiologists. We address this issue with a convolutional neural network (CNN)-based technique, synthesiZed Improved Resolution and Concurrent nOise reductioN (ZIRCON), that creates a single, thin, low-noise series that combines the favorable features from smooth and sharp head kernels. ZIRCON uses a CNN model with an autoencoder U-Net architecture that accepts two input channels (smooth- and sharp-kernel CT images) and combines their salient features to produce a single CT image. Image quality requirements are built into a task-based loss function with a smooth and sharp loss terms specific to anatomical regions. The model is trained using supervised learning with paired routine-dose clinical non-contrast head CT images as training targets and simulated low-dose (25%) images as training inputs. One hundred unique de-identified clinical exams were used for training, ten for validation, and ten for testing. Visual comparisons and contrast measurements of ZIRCON revealed that thinner slices and the smooth-kernel loss function improved gray-white matter contrast. Combined with lower noise, this increased visibility of small soft-tissue features that would be otherwise impaired by partial volume averaging or noise. Line profile analysis showed that ZIRCON images largely retained sharpness compared to the sharp-kernel input images. ZIRCON combined desirable image quality properties of both smooth and sharp input kernels into a single, thin, low-noise series suitable for both brain and skull imaging.

Image quality evaluation in deep-learning-based CT noise reduction using virtual imaging trial methods: Contrast-dependent spatial resolution

BACKGROUND

Deep-learning-based image reconstruction and noise reduction methods (DLIR) have been increasingly deployed in clinical CT. Accurate image quality assessment of these methods is challenging as the performance measured using physical phantoms may not represent the true performance of DLIR in patients since DLIR is trained mostly on patient images.

PURPOSE

In this work, we aim to develop a patient-data-based virtual imaging trial framework and, as a first application, use it to measure the spatial resolution properties of a DLIR method.

METHODS

The patient-data-based virtual imaging trial framework consists of five steps: (1) insertion of lesions into projection domain data using the acquisition geometry of the patient exam to simulate different lesion characteristics; (2) insertion of noise into projection domain data using a realistic photon statistical model of the CT system to simulate different dose levels; (3) creation of DLIR-processed images from projection or image data; (4) creation of ensembles of DLIR-processed patient images from a large number of noise and lesion realizations; and (5) evaluation of image quality using ensemble DLIR images. This framework was applied to measure the spatial resolution of a ResNet based deep convolutional neural network (DCNN) trained on patient images. Lesions in a cylindrical shape and different contrast levels (-500, -100, -50, -20, -10 HU) were inserted to the lower right lobe of the liver in a patient case. Multiple dose levels were simulated (50%, 25%, 12.5%). Each lesion and dose condition had 600 noise realizations. Multiple reconstruction and denoising methods were used on all the noise realizations, including the original filtered-backprojection (FBP), iterative reconstruction (IR), and the DCNN method with three different strength setting (DCNN-weak, DCNN-medium, and DCNN-strong). Mean lesion signal was calculated by performing ensemble averaging of all the noise realizations for each lesion and dose condition and then subtracting the lesion-present images from the lesion absent images. Modulation transfer functions (MTFs) both in-plane and along the z-axis were calculated based on the mean lesion signals. The standard deviations of MTFs at each condition were estimated with bootstrapping: randomly sampling (with replacement) all the DLIR/FBP/IR images from the ensemble data (600 samples) at each condition. The impact of varying lesion contrast, dose levels, and denoising strengths were evaluated. Statistical analysis with paired t-test was used to compare the z-axis and in-plane spatial resolution of five algorithms for five different contrasts and three dose levels.

RESULTS

The in-plane and z-axis spatial resolution degradation of DCNN becomes more severe as the contrast or radiation dose decreased, or DCNN denoising strength increased. In comparison with FBP, a 59.5% and 4.1% reduction of in-plane and z-axis MTF (in terms of spatial frequencies at 50% MTF), respectively, was observed at low contrast (-10 HU) for DCNN with the highest denoising strength at 25% routine dose level. When the dose level reduces from 50% to 12.5% of routine dose, the in-plane and z-axis MTFs reduces from 92.1% to 76.3%, and from 98.9% to 95.5%, respectively, at contrast of -100 HU, using FBP as the reference. For most conditions of contrasts and dose levels, significant differences were found among the five algorithms, with the following relationship in both in-plane and cross-plane spatial resolution: FBP > DCNN-Weak > IR > DCNN-Medium > DCNN-Strong. The spatial resolution difference among algorithms decreases at higher contrast or dose levels.

CONCLUSIONS

A patient-data-based virtual imaging trial framework was developed and applied to measuring the spatial resolution properties of a DCNN noise reduction method at different contrast and dose levels using real patient data. As with other non-linear image reconstruction and post-processing techniques, the evaluated DCNN method degraded the in-plane and z-axis spatial resolution at lower contrast levels, lower radiation dose, and higher denoising strength.

Pediatric evaluations for deep learning CT denoising

BACKGROUND

Deep learning (DL) CT denoising models have the potential to improve image quality for lower radiation dose exams. These models are generally trained with large quantities of adult patient image data. However, CT, and increasingly DL denoising methods, are used in both adult and pediatric populations. Pediatric body habitus and size can differ significantly from adults and vary dramatically from newborns to adolescents. Ensuring that pediatric subgroups of different body sizes are not disadvantaged by DL methods requires evaluations capable of assessing performance in each subgroup.

PURPOSE

To assess DL CT denoising in pediatric and adult-sized patients, we built a framework of computer simulated image quality (IQ) control phantoms and evaluation methodology.

METHODS

The computer simulated IQ phantoms in the framework featured pediatric-sized versions of standard CatPhan 600 and MITA-LCD phantoms with a range of diameters matching the mean effective diameters of pediatric patients ranging from newborns to 18 years old. These phantoms were used in simulating CT images that were then inputs for a DL denoiser to evaluate performance in different sized patients. Adult CT test images were simulated using standard-sized phantoms scanned with adult scan protocols. Pediatric CT test images were simulated with pediatric-sized phantoms and adjusted pediatric protocols. The framework's evaluation methodology consisted of denoising both adult and pediatric test images then assessing changes in image quality, including noise, image sharpness, CT number accuracy, and low contrast detectability. To demonstrate the use of the framework, a REDCNN denoising model trained on adult patient images was evaluated. To validate that the DL model performance measured with the proposed pediatric IQ phantoms was representative of performance in more realistic patient anatomy, anthropomorphic pediatric XCAT phantoms of the same age range were also used to compare noise reduction performance.

RESULTS

Using the proposed pediatric-sized IQ phantom framework, size differences between adult and pediatric-sized phantoms were observed to substantially influence the adult trained DL denoising model's performance. When applied to adult images, the DL model achieved a 60% reduction in noise standard deviation without substantial loss in sharpness in mid or high spatial frequencies. However, in smaller phantoms the denoising performance dropped due to different image noise textures resulting from the smaller field of view (FOV) between adult and pediatric protocols. In the validation study, noise reduction trends in the pediatric-sized IQ phantoms were found to be consistent with those found in anthropomorphic phantoms.

CONCLUSION

We developed a framework of using pediatric-sized IQ phantoms for pediatric subgroup evaluation of DL denoising models. Using the framework, we found the performance of an adult trained DL denoiser did not generalize well in the smaller diameter phantoms corresponding to younger pediatric patient sizes. Our work suggests noise texture differences from FOV changes between adult and pediatric protocols can contribute to poor generalizability in DL denoising and that the proposed framework is an effective means to identify these performance disparities for a given model.

AntiHalluciNet: A Potential Auditing Tool of the Behavior of Deep Learning Denoising Models in Low-Dose Computed Tomography

Gaining the ability to audit the behavior of deep learning (DL) denoising models is of crucial importance to prevent potential hallucinations and adversarial clinical consequences. We present a preliminary version of AntiHalluciNet, which is designed to predict spurious structural components embedded in the residual noise from DL denoising models in low-dose CT and assess its feasibility for auditing the behavior of DL denoising models. We created a paired set of structure-embedded and pure noise images and trained AntiHalluciNet to predict spurious structures in the structure-embedded noise images. The performance of AntiHalluciNet was evaluated by using a newly devised residual structure index (RSI), which represents the prediction confidence based on the presence of structural components in the residual noise image. We also evaluated whether AntiHalluciNet could assess the image fidelity of a denoised image by using only a noise component instead of measuring the SSIM, which requires both reference and test images. Then, we explored the potential of AntiHalluciNet for auditing the behavior of DL denoising models. AntiHalluciNet was applied to three DL denoising models (two pre-trained models, RED-CNN and CTformer, and a commercial software, ClariCT.AI [version 1.2.3]), and whether AntiHalluciNet could discriminate between the noise purity performances of DL denoising models was assessed. AntiHalluciNet demonstrated an excellent performance in predicting the presence of structural components. The RSI values for the structural-embedded and pure noise images measured using the 50% low-dose dataset were 0.57 ± 31 and 0.02 ± 0.02, respectively, showing a substantial difference with a p-value < 0.0001. The AntiHalluciNet-derived RSI could differentiate between the quality of the degraded denoised images, with measurement values of 0.27, 0.41, 0.48, and 0.52 for the 25%, 50%, 75%, and 100% mixing rates of the degradation component, which showed a higher differentiation potential compared with the SSIM values of 0.9603, 0.9579, 0.9490, and 0.9333. The RSI measurements from the residual images of the three DL denoising models showed a distinct distribution, being 0.28 ± 0.06, 0.21 ± 0.06, and 0.15 ± 0.03 for RED-CNN, CTformer, and ClariCT.AI, respectively. AntiHalluciNet has the potential to predict the structural components embedded in the residual noise from DL denoising models in low-dose CT. With AntiHalluciNet, it is feasible to audit the performance and behavior of DL denoising models in clinical environments where only residual noise images are available.

Supervised Learning for CT Denoising and Deconvolution Without High-Resolution Reference Images

Purpose

Convolutional neural networks (CNNs) have been proposed for super-resolution in CT, but training of CNNs requires high-resolution reference data. Higher spatial resolution can also be achieved using deconvolution, but conventional deconvolution approaches amplify noise. We develop a CNN that mitigates increasing noise and that does not require higher-resolution reference images.

Methods

Our model includes a noise reduction CNN and a deconvolution CNN that are separately trained. The noise reduction CNN is a U-Net, similar to other noise reduction CNNs found in the literature. The deconvolution CNN uses an autoencoder, where the decoder is fixed and provided as a hyperparameter that represents the system point spread function. The encoder is trained to provide a deconvolution that does not amplify noise. Ringing can occur from deconvolution but is controlled with a difference of gradients loss function term. Our technique was demonstrated on a variety of patient images and on ex vivo kidney stones.

Results

The noise reduction and deconvolution CNNs produced visually sharper images at low noise. In ex vivo mixed kidney stones, better visual delineation of the kidney stone components could be seen.

Conclusions

A noise reduction and deconvolution CNN improves spatial resolution and reduces noise without requiring higher-resolution reference images.

Self-trained deep convolutional neural network for noise reduction in CT

Purpose

Supervised deep convolutional neural network (CNN)-based methods have been actively used in clinical CT to reduce image noise. The networks of these methods are typically trained using paired high- and low-quality data from a massive number of patients and/or phantom images. This training process is tedious, and the network trained under a given condition may not be generalizable to patient images acquired and reconstructed under different conditions. We propose a self-trained deep CNN (ST_CNN) method for noise reduction in CT that does not rely on pre-existing training datasets.

Approach

The ST_CNN training was accomplished using extensive data augmentation in the projection domain, and the inference was applied to the data itself. Specifically, multiple independent noise insertions were applied to the original patient projection data to generate multiple realizations of low-quality projection data. Then, rotation augmentation was adopted for both the original and low-quality projection data by applying the rotation angle directly on the projection data so that images were rotated at arbitrary angles without introducing additional bias. A large number of paired low- and high-quality images from the same patient were reconstructed and paired for training the ST_CNN model.

Results

No significant difference was found between the ST_CNN and conventional CNN models in terms of the peak signal-to-noise ratio and structural similarity index measure. The ST_CNN model outperformed the conventional CNN model in terms of noise texture and homogeneity in liver parenchyma as well as better subjective visualization of liver lesions. The ST_CNN may sacrifice the sharpness of vessels slightly compared to the conventional CNN model but without affecting the visibility of peripheral vessels or diagnosis of vascular pathology.

Conclusions

The proposed ST_CNN method trained from the data itself may achieve similar image quality in comparison with conventional deep CNN denoising methods pre-trained on external datasets.

Deep Learning-Based Image Noise Quantification Framework for Computed Tomography

OBJECTIVE

Noise quantification is fundamental to computed tomography (CT) image quality assessment and protocol optimization. This study proposes a deep learning-based framework, Single-scan Image Local Variance EstimatoR (SILVER), for estimating the local noise level within each region of a CT image. The local noise level will be referred to as a pixel-wise noise map.

METHODS

The SILVER architecture resembled a U-Net convolutional neural network with mean-square-error loss. To generate training data, 100 replicate scans were acquired of 3 anthropomorphic phantoms (chest, head, and pelvis) using a sequential scan mode; 120,000 phantom images were allocated into training, validation, and testing data sets. Pixel-wise noise maps were calculated for the phantom data by taking the per-pixel SD from the 100 replicate scans. For training, the convolutional neural network inputs consisted of phantom CT image patches, and the training targets consisted of the corresponding calculated pixel-wise noise maps. Following training, SILVER noise maps were evaluated using phantom and patient images. For evaluation on patient images, SILVER noise maps were compared with manual noise measurements at the heart, aorta, liver, spleen, and fat.

RESULTS

When tested on phantom images, the SILVER noise map prediction closely matched the calculated noise map target (root mean square error <8 Hounsfield units). Within 10 patient examinations, SILVER noise map had an average percent error of 5% relative to manual region-of-interest measurements.

CONCLUSION

The SILVER framework enabled accurate pixel-wise noise level estimation directly from patient images. This method is widely accessible because it operates in the image domain and requires only phantom data for training.

Non-contrast CT synthesis using patch-based cycle-consistent generative adversarial network (Cycle-GAN) for radiomics and deep learning in the era of COVID-19

Handcrafted and deep learning (DL) radiomics are popular techniques used to develop computed tomography (CT) imaging-based artificial intelligence models for COVID-19 research. However, contrast heterogeneity from real-world datasets may impair model performance. Contrast-homogenous datasets present a potential solution. We developed a 3D patch-based cycle-consistent generative adversarial network (cycle-GAN) to synthesize non-contrast images from contrast CTs, as a data homogenization tool. We used a multi-centre dataset of 2078 scans from 1,650 patients with COVID-19. Few studies have previously evaluated GAN-generated images with handcrafted radiomics, DL and human assessment tasks. We evaluated the performance of our cycle-GAN with these three approaches. In a modified Turing-test, human experts identified synthetic vs acquired images, with a false positive rate of 67% and Fleiss' Kappa 0.06, attesting to the photorealism of the synthetic images. However, on testing performance of machine learning classifiers with radiomic features, performance decreased with use of synthetic images. Marked percentage difference was noted in feature values between pre- and post-GAN non-contrast images. With DL classification, deterioration in performance was observed with synthetic images. Our results show that whilst GANs can produce images sufficient to pass human assessment, caution is advised before GAN-synthesized images are used in medical imaging applications.

Structure-preserved meta-learning uniting network for improving low-dose CT quality

Objective.Deep neural network (DNN) based methods have shown promising performances for low-dose computed tomography (LDCT) imaging. However, most of the DNN-based methods are trained on simulated labeled datasets, and the low-dose simulation algorithms are usually designed based on simple statistical models which deviate from the real clinical scenarios, which could lead to issues of overfitting, instability and poor robustness. To address these issues, in this work, we present a structure-preserved meta-learning uniting network (shorten as 'SMU-Net') to suppress noise-induced artifacts and preserve structure details in the unlabeled LDCT imaging task in real scenarios.Approach.Specifically, the presented SMU-Net contains two networks, i.e., teacher network and student network. The teacher network is trained on simulated labeled dataset and then helps the student network train with the unlabeled LDCT images via the meta-learning strategy. The student network is trained on real LDCT dataset with the pseudo-labels generated by the teacher network. Moreover, the student network adopts the Co-teaching strategy to improve the robustness of the presented SMU-Net.Main results.We validate the proposed SMU-Net method on three public datasets and one real low-dose dataset. The visual image results indicate that the proposed SMU-Net has superior performance on reducing noise-induced artifacts and preserving structure details. And the quantitative results exhibit that the presented SMU-Net method generally obtains the highest signal-to-noise ratio (PSNR), the highest structural similarity index measurement (SSIM), and the lowest root-mean-square error (RMSE) values or the lowest natural image quality evaluator (NIQE) scores.Significance.We propose a meta learning strategy to obtain high-quality CT images in the LDCT imaging task, which is designed to take advantage of unlabeled CT images to promote the reconstruction performance in the LDCT environments.