Radiomics is attracting research interests for characterization of the tumor phenotype as well as for prediction of patient outcome. However, many radiomic features are known to be affected by a multitude of variability sources, such as CT acquisition parameters, which might lead to false discovery if unknowingly used. Therefore, in order to avoid such pitfalls, the appropriate selection of robust features is an essential task in radiomic studies. We investigate the variability of CT imaging features which were previously reported as radiomic markers in non- small cell lung cancer (NSCLC). We scanned a standardized phantom with 64-slice multi-detector CT scanner with various scan conditions. We extracted forty-seven radiomic features including two texture features and first order statistics. Feature variability index was measured to evaluate the feature robustness depending on the scan parameters. The proportion of feature less effect on kernel was observed to only 32%. Our study revealed a high variability of CT image features depending on technical parameters. These characteristics should be considered in the feature extraction procedure when different protocols are used in the patient dataset. Use of the same CT protocol is preferred. Otherwise, the application of kernel normalization techniques is necessary for the radiomic study.
This study presents a novel deep learning approach for denoising of ultra-low-dose cardiac CT angiography (CCTA) by combining a low-dose simulation technique and convolutional neural network (CNN). Twenty-five CT angiography (CTA) scans acquired with ECG gating (70 – 100 kVp, 100 – 200 mAs) were fed into the low-dose simulation tool to generate a paired set of simulated low-dose CTA and synthetic low-dose noise. A modified U-net model with 4x4 kernel size and five layers was trained with these paired dataset to predict the low-dose noise from the given low-dose CCTA image. For generation of simulation low-dose CTA, differing level of low-dose conditions from 10% to 2.5% were applied. Independent 5 ultra-low-dose CTA scans (70 – 100 kVp, 4% dose of full-dose) with ECG gating were used for testing the denoising performance of the trained U-net. A denoised CCTA image was obtained by subtracting the predicted noise image by the U-net from the ultra-low-dose CCTA images. The performance was evaluated quantitatively in terms of noise measurements in ascending aorta, left/right ventricles, and qualitatively by comparing the noise pattern and image quality. Average of image noise in ascending aorta, left/right ventricles were 149±41HU, 200±15HU, 164±21HU in ultra-low-dose, and 46±14HU, 66±9HU, 55±12HU in deep learning-denoised images. The overall noise was significantly reduced by 70%. The noise pattern was indistinguishable from that of real CCTA image, and the image quality of denoised CCTA images was much higher than that of ultra-lowdose CCTA images.
Effective elimination of unique CT noise pattern while preserving adequate image quality is crucial in reducing radiation dose to ultra-low-dose level in CT imaging practice. In this study, we present a novel Deep Learning-enable Iterative Reconstruction (Deep IR) approach for CT denoising which incorporate a synthetic sinogram-based noise simulation technique for training of Convolutional Neural Network (CNN). Regular dose CT images from 25 patients were used from Seoul National University Hospital. The CT scans were performed at 140 kVp, 100 mAs, and reconstructed with standard FBP technique using B60f kernel. Among them, 20 patients were randomly selected as a training set and the rest 5 patients were used for a test set. We applied a re-projection technique to create a synthetic sinogram from the DICOM CT image, and then a simulated noise sinogram was generated to match the noise level of 10mAs according to Poisson statistic and the system noise model of the given scanner (Somatom Sensation 16, Siemens). We added the simulated noise sinogram to the re-projected synthetic sinogram to generate a simulated sinogram of ultra-low dose scan. We also created the simulated ultra-low-dose CT image by applying FBP reconstruction of the simulated noise sinogram with B60f kernel. A CNN model was created using a TensorFlow framework to have 10 consecutive convolution layer and activation layer. The CNN was trained to learn the noise in sinogram domain: the simulated noisy sinogram of ultra-low dose scan was fed into its input nodes with the output node being fed by the simulated noise sinogram. At test phase, the noise sinogram from the CNN output was reconstructed with using B60f kernel to create a noise CT image, which in turn was subtracted from the simulated ultra-low-dose CT image to produce a Deep IR CT image. The performance was evaluated quantitatively in terms of structural similarity (SSIM) index, peak signal-to-noise ratio (PSNR) and noise level measurement and qualitatively in CT image by comparing the noise pattern and image quality. Compared to low-dose image, denoising image of the SSIM and the PSNR were improved from 0.75 to 0.80, 28.61db to 32.16 respectively. The noise level of denoising image was reduced to an average of 56 % of that of low-dose image. The noise pattern in reconstructed noise CT was indistinguishable from that of real CT images, and the image quality of Deep IR CT image was overall much higher than that of simulated ultra-low-dose CT.
Differing reconstruction kernels are known to strongly affect the variability of imaging biomarkers and thus remain as a barrier in translating the computer aided quantification techniques into clinical practice. This study presents a deep learning application to CT kernel conversion which converts a CT image of sharp kernel to that of standard kernel and evaluates its impact on variability reduction of a pulmonary imaging biomarker, the emphysema index (EI). Forty cases of low-dose chest CT exams obtained with 120kVp, 40mAs, 1mm thickness, of 2 reconstruction kernels (B30f, B50f) were selected from the low dose lung cancer screening database of our institution. A Fully convolutional network was implemented with Keras deep learning library. The model consisted of symmetric layers to capture the context and fine structure characteristics of CT images from the standard and sharp reconstruction kernels. Pairs of the full-resolution CT data set were fed to input and output nodes to train the convolutional network to learn the appropriate filter kernels for converting the CT images of sharp kernel to standard kernel with a criterion of measuring the mean squared error between the input and target images. EIs (RA950 and Perc15) were measured with a software package (ImagePrism Pulmo, Seoul, South Korea) and compared for the data sets of B50f, B30f, and the converted B50f. The effect of kernel conversion was evaluated with the mean and standard deviation of pair-wise differences in EI. The population mean of RA950 was 27.65 ± 7.28% for B50f data set, 10.82 ± 6.71% for the B30f data set, and 8.87 ± 6.20% for the converted B50f data set. The mean of pair-wise absolute differences in RA950 between B30f and B50f is reduced from 16.83% to 1.95% using kernel conversion. Our study demonstrates the feasibility of applying the deep learning technique for CT kernel conversion and reducing the kernel-induced variability of EI quantification. The deep learning model has a potential to improve the reliability of imaging biomarker, especially in evaluating the longitudinal changes of EI even when the patient CT scans were performed with different kernels.
Mammographic breast density is a well-established marker for breast cancer risk. However, accurate measurement of dense tissue is a difficult task due to faint contrast and significant variations in background fatty tissue. This study presents a novel method for automated mammographic density estimation based on Convolutional Neural Network (CNN). A total of 397 full-field digital mammograms were selected from Seoul National University Hospital. Among them, 297 mammograms were randomly selected as a training set and the rest 100 mammograms were used for a test set. We designed a CNN architecture suitable to learn the imaging characteristic from a multitudes of sub-images and classify them into dense and fatty tissues. To train the CNN, not only local statistics but also global statistics extracted from an image set were used. The image set was composed of original mammogram and eigen-image which was able to capture the X-ray characteristics in despite of the fact that CNN is well known to effectively extract features on original image. The 100 test images which was not used in training the CNN was used to validate the performance. The correlation coefficient between the breast estimates by the CNN and those by the expert’s manual measurement was 0.96. Our study demonstrated the feasibility of incorporating the deep learning technology into radiology practice, especially for breast density estimation. The proposed method has a potential to be used as an automated and quantitative assessment tool for mammographic breast density in routine practice.
Iterative reconstruction (IR) technique is growingly used in clinical CT imaging with an expectation for improved image
quality at lower patient doses. However, the nonlinear frequency response in different noise level and object contrast is
less explored. In this study, we evaluate object contrast and dose level-dependent behavior of modulation transfer
function in iterative reconstruction computed tomography imaging with a specially fabricated phantom. We created
multi-contrast edge phantom, which consists of acrylic panel and diluted iodine contrast agent with different
concentrations. Images were acquired with a multi-detector CT (Discovery CT750 HD: GE) at four dose levels (25, 50,
100 and 200mAs), and were reconstructed using FBP and two IR techniques (ASIR50 and VEO). Edge spread functions
were extracted across angled edges on image, and were differentiated to yield line spread function. LSF were Fourier
transformed to evaluate the presampled MTFs of IR and FBP reconstruction techniques. At same dose level (200mAs),
the MTFs with higher contrast showed higher response than that of lower contrast in VEO. A MTF50 of 200mAs showed
markedly higher responses up to 23% than that of 25mAs scan with VEO reconstruction for an edge phantom of 520HU
contrast. Our study revealed that MTF of IR technique degrades depending on noise level at low dose scan. Therefore,
we recommend that its characteristic should be considered in quantitative analysis such as lesion size measurement.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.