Journal: Journal of Imaging

Article Title: Low-Cost Probabilistic 3D Denoising with Applications for Ultra-Low-Radiation Computed Tomography

doi: 10.3390/jimaging8060156

Figure Lengend Snippet: Graphical representation of our proposed pipeline workflow for automated generation and risk assessment of CT images. ( A ): Initial reference data can be either a set of real CT-data generated using high-dose radiation or artificially simulated data. ( B ) Exemplary high-quality and low-quantum noise image of lung vessels. ( C ) exemplary low-dose CT images with high quantum noise. ( D ) Workflow from image generation to subsequent benchmarking of ML/DL-denoising methods. Starting with high-quality data or artificially generated reference data, respectively, a spectrum of image noise σ is added for a multitude of combinations from patient-specific and CT control variables, as suggested in Equation . The noisy images are then denoised using various state-of-the-art methods and the processed images are compared to the original reference data.

Article Snippet: We used denoising methods based on local window filtering of the data (3D Gaussian filtering with the MATLAB function imgaussfilt3() , 3D local median filtering with the MATLAB function medfilt3() and bilateral filtering with the MATLAB function imbilatfilt() ) [ , , , ], spectral denoising methods (the 3D wavelets denoising with the MATLAB function wavedec3() ) [ , , , , ] and a deep learning denoising method based on pre-trained feed-forward denoising convolutional neural networks (DnCNNs, with the MATLAB functions denoiseImage() and denoisingNetwork() ) [ , , , ].

Techniques: Generated, Control

Journal: Journal of Imaging

Article Title: Low-Cost Probabilistic 3D Denoising with Applications for Ultra-Low-Radiation Computed Tomography

doi: 10.3390/jimaging8060156

Figure Lengend Snippet: Graphical overview of the Probabilistic Mumford–Shah (PMS) framework. ( A ) Summary of the parameters and variables. ( B ) Core rSPA algorithm idea: 3D-denoising with the regularized Scalable Probabilistic Approximation algorithm (rSPA). Given the (noisy) CT voxel data V , rSPA minimizes the function L ( C , Γ ) and seeks for the optimal segmentation of V in terms of the K spatially-persistent latent features characterized by the latent feature probabilities in K rows of the matrix Γ , as well as by the latent colors as K columns of the latent color matrix C . Persistency of the feature segmentation is imposed by the second term of the right-hand side of the function L ( C , Γ ) , which penalizes the differences in feature probability values in the spatially-neighboring points. ( C ) Denoising idea: latent feature probabilities are persistent (slowly-changing) 3D functions. ( D ) Graphical representation of the overlapping domain decomposition used in the parallel DD-rSPA algorithm.

Article Snippet: We used denoising methods based on local window filtering of the data (3D Gaussian filtering with the MATLAB function imgaussfilt3() , 3D local median filtering with the MATLAB function medfilt3() and bilateral filtering with the MATLAB function imbilatfilt() ) [ , , , ], spectral denoising methods (the 3D wavelets denoising with the MATLAB function wavedec3() ) [ , , , , ] and a deep learning denoising method based on pre-trained feed-forward denoising convolutional neural networks (DnCNNs, with the MATLAB functions denoiseImage() and denoisingNetwork() ) [ , , , ].

Techniques:

Journal: Journal of Imaging

Article Title: Low-Cost Probabilistic 3D Denoising with Applications for Ultra-Low-Radiation Computed Tomography

doi: 10.3390/jimaging8060156

Figure Lengend Snippet: Radiation exposure, quantum noise, and denoising performance of CNNs and rSPA in low-radiation and ultra-low-radiation thorax CT regimes. ( A ) Reference data of a thorax CT voxel fragment (approx. 5 cm 3 ) of a 19-year-old female with the BMI 27.5, acquired with the Somatum Emotion 16 2007 (Siemens Aktiengesellschaft, Berlin, Germany) at 130 kV tube voltage. ( B ) Simulated decrease in the radiation exposure CTDI vol from 15.6 mGy (reference frame) to 3.3 mGy (for low-radiation simulations) and 0.5 mGy (ultra-low-radiation) results in a significant increase of quantum noise. ( C ) Reconstructed images using CNNs. ( D ) Reconstructed images using rSPA. ( E ) 3D segmentation of the original reference frame. ( F ) 3D segmentation based on the images denoised using CNNs. ( G ) 3D-segmentation of the images denoised by rSPA.

Article Snippet: We used denoising methods based on local window filtering of the data (3D Gaussian filtering with the MATLAB function imgaussfilt3() , 3D local median filtering with the MATLAB function medfilt3() and bilateral filtering with the MATLAB function imbilatfilt() ) [ , , , ], spectral denoising methods (the 3D wavelets denoising with the MATLAB function wavedec3() ) [ , , , , ] and a deep learning denoising method based on pre-trained feed-forward denoising convolutional neural networks (DnCNNs, with the MATLAB functions denoiseImage() and denoisingNetwork() ) [ , , , ].

Techniques:

Journal: Journal of Imaging

Article Title: Low-Cost Probabilistic 3D Denoising with Applications for Ultra-Low-Radiation Computed Tomography

doi: 10.3390/jimaging8060156

Figure Lengend Snippet: Comparing denoising performance on synthetic CT images of noisy circles, with DL from additionally trained to recognize circles for non-Gaussian noise model: ( A ) medium noise scenario, corresponding to low-radiation regime with around 3.3 mGy; ( B ) high noise scenario, corresponding to ultra-low-radiation regime with 0.5 mGy.

Article Snippet: We used denoising methods based on local window filtering of the data (3D Gaussian filtering with the MATLAB function imgaussfilt3() , 3D local median filtering with the MATLAB function medfilt3() and bilateral filtering with the MATLAB function imbilatfilt() ) [ , , , ], spectral denoising methods (the 3D wavelets denoising with the MATLAB function wavedec3() ) [ , , , , ] and a deep learning denoising method based on pre-trained feed-forward denoising convolutional neural networks (DnCNNs, with the MATLAB functions denoiseImage() and denoisingNetwork() ) [ , , , ].

Techniques:

Journal: Journal of Imaging

Article Title: Low-Cost Probabilistic 3D Denoising with Applications for Ultra-Low-Radiation Computed Tomography

doi: 10.3390/jimaging8060156

Figure Lengend Snippet: Comparing denoising quality, cost and parallelizability: ( A – C ) comparison of PMS rSPA algorithm to the regularized Mumford–Shah denoising tool introduced in and to the additionally trained DL denoising algorithm from and ; ( D , E ) computational cost scaling and performance for DL (without taking into account time for additional training), sequential rSPA, parallel DD-rSPA and DD-rSPA followed by DL. Each point of each method’s curve and surface is obtained from statistical averaging of the respective values obtained by analyzing 10 randomly-generated images with these particular combinations of image size and noise level.

Article Snippet: We used denoising methods based on local window filtering of the data (3D Gaussian filtering with the MATLAB function imgaussfilt3() , 3D local median filtering with the MATLAB function medfilt3() and bilateral filtering with the MATLAB function imbilatfilt() ) [ , , , ], spectral denoising methods (the 3D wavelets denoising with the MATLAB function wavedec3() ) [ , , , , ] and a deep learning denoising method based on pre-trained feed-forward denoising convolutional neural networks (DnCNNs, with the MATLAB functions denoiseImage() and denoisingNetwork() ) [ , , , ].

Techniques: Comparison, Generated

Journal: Journal of Imaging

Article Title: Low-Cost Probabilistic 3D Denoising with Applications for Ultra-Low-Radiation Computed Tomography

doi: 10.3390/jimaging8060156

Figure Lengend Snippet: Comparing CT image denoising performances for three CT noise models: ( A ) additive Gaussian noise model (CT noise variance is independent of the feature color); ( B ) multiplicative non-Gaussian noise model (CT noise variance changes with the amplitude of the underlying color signal); ( C ) empirical noise obtained from the thorax CT patient data. In ( A , B ), generation of synthetic images was performed for a patient with a water-equivalent diameter of 30 cm, which is subject to a Thorax CT with a typical tube voltage of 120 kV in the range of tube currents between 5–180 mA and a set of artificial anatomic features from A (with a feature contrast of 200 HU). In ( C ), real patient data were used. Comparison is performed with three primary image quality criteria: with mean squared error (left panels); with peak signal-to-noise ratio (middle panels); and with the 3D multiscale structural similarity index (right panels).

Article Snippet: We used denoising methods based on local window filtering of the data (3D Gaussian filtering with the MATLAB function imgaussfilt3() , 3D local median filtering with the MATLAB function medfilt3() and bilateral filtering with the MATLAB function imbilatfilt() ) [ , , , ], spectral denoising methods (the 3D wavelets denoising with the MATLAB function wavedec3() ) [ , , , , ] and a deep learning denoising method based on pre-trained feed-forward denoising convolutional neural networks (DnCNNs, with the MATLAB functions denoiseImage() and denoisingNetwork() ) [ , , , ].

Techniques: Comparison

Journal: Journal of Imaging

Article Title: Low-Cost Probabilistic 3D Denoising with Applications for Ultra-Low-Radiation Computed Tomography

doi: 10.3390/jimaging8060156

Figure Lengend Snippet: Comparing denoising methods with the average Multiscale Structural Similarity Index (3D MS-SSIM): ( A ) varying the true underlying feature contrast and LAR for a synthetic 30-year-old female patient with a water-equiv. cross-section of 27 cm; ( B ) varying the true underlying feature contrast and LAR for a synthetic 1-year-old female infant patient with a water-equiv. cross-section of 12.7 cm; ( C ) denoising performance comparison when varying the patient size and the effective absorbed radiation dose density, with the 200 Hounsfield Units (HU) feature contrast differences.

Article Snippet: We used denoising methods based on local window filtering of the data (3D Gaussian filtering with the MATLAB function imgaussfilt3() , 3D local median filtering with the MATLAB function medfilt3() and bilateral filtering with the MATLAB function imbilatfilt() ) [ , , , ], spectral denoising methods (the 3D wavelets denoising with the MATLAB function wavedec3() ) [ , , , , ] and a deep learning denoising method based on pre-trained feed-forward denoising convolutional neural networks (DnCNNs, with the MATLAB functions denoiseImage() and denoisingNetwork() ) [ , , , ].

Techniques: Comparison

Journal: Journal of Imaging

Article Title: Low-Cost Probabilistic 3D Denoising with Applications for Ultra-Low-Radiation Computed Tomography

doi: 10.3390/jimaging8060156

Figure Lengend Snippet: Deterioration of CT image quality (decrease in 3D MS-SSIM index, baseline = 100%) caused by a reduction of lifetime attributable risk (LAR) for different methods. The CT scans pertain to the infant patient, with a fixed feature contrast of 200 HU.

Article Snippet: We used denoising methods based on local window filtering of the data (3D Gaussian filtering with the MATLAB function imgaussfilt3() , 3D local median filtering with the MATLAB function medfilt3() and bilateral filtering with the MATLAB function imbilatfilt() ) [ , , , ], spectral denoising methods (the 3D wavelets denoising with the MATLAB function wavedec3() ) [ , , , , ] and a deep learning denoising method based on pre-trained feed-forward denoising convolutional neural networks (DnCNNs, with the MATLAB functions denoiseImage() and denoisingNetwork() ) [ , , , ].

Techniques:

Journal: Journal of Imaging

Article Title: Low-Cost Probabilistic 3D Denoising with Applications for Ultra-Low-Radiation Computed Tomography

doi: 10.3390/jimaging8060156

Figure Lengend Snippet: Comparing denoising methods with the average Multiscale Structural Similarity Index (3D MS-SSIM) for simulated thorax CT: ( A ) varying the absorbed radiation dose for a synthetic 30-year-old female patient with a water-equiv. cross-section of 27 cm; ( B ) varying the absorbed radiation dose for a synthetic 1-year-old female infant patient with a water-equiv. cross-section of 12.7 cm. Noiseless thorax CT image used as reference in this performance comparison is available at https://www.dropbox.com/s/29x0xivg8l80q10/female_lung_thorax_CT_image_section_v2.mat?dl=0 (accessed on 18 March 2022). Dotted lines show 95% nonparametric confidence intervals (c.i.) obtained for every value of CTDI vol from 100 different independently-generated noisy synthetic CT images, using the MATLAB function quantile() .

Article Snippet: We used denoising methods based on local window filtering of the data (3D Gaussian filtering with the MATLAB function imgaussfilt3() , 3D local median filtering with the MATLAB function medfilt3() and bilateral filtering with the MATLAB function imbilatfilt() ) [ , , , ], spectral denoising methods (the 3D wavelets denoising with the MATLAB function wavedec3() ) [ , , , , ] and a deep learning denoising method based on pre-trained feed-forward denoising convolutional neural networks (DnCNNs, with the MATLAB functions denoiseImage() and denoisingNetwork() ) [ , , , ].

Techniques: Comparison, Generated