Denoising Tc-99m DMSA images using Denoising Convolutional Neural Network with comparison to a Block Matching Filter.

INTRODUCTION: A DnCNN for image denoising trained with natural images is available in MATLAB. For Tc-99m DMSA images, any loss of clinical details during the denoising process will have serious consequences since denoised image is to be used for diagnosis. The objective of the study was to find whether this pre-trained DnCNN can be used for denoising Tc-99m DMSA images and compare its performance with block matching 3D (BM3D) filter. MATERIALS AND METHODS: Two hundred forty-two Tc-99m DMSA images were denoised using BM3D filter (at sigma = 5, 10, 15, 20, and 25) and DnCNN. The original and denoised images were reviewed by two nuclear medicine physicians and also assessed objectively using the image quality metrics: SSIM, FSIM, MultiSSIM, PIQE, Blur, GCF, and Brightness. Wilcoxon signed-rank test was applied to find the statistically significant difference between the value of image quality metrics of the denoised images and the corresponding original images. RESULTS: Nuclear medicine physicians observed no loss of clinical information in DnCNN denoised image and superior image quality compared to its original and BM3D denoised images. Edges/boundaries of the scar were found to be well preserved, and doubtful scar became obvious in the denoised image. Objective assessment also showed that the quality of DnCNN denoised images was significantly better than that of original images at P -value <0.0001. CONCLUSION: The pre-trained DnCNN available with MATLAB Deep Learning Toolbox can be used for denoising Tc-99m DMSA images, and the performance of DnCNN was found to be superior in comparison with BM3D filter.

Vertebral and ribs enhancement in Tc-99m methylene diphosphonate bone scan image by blind deconvolution algorithm.

AIMS AND OBJECTIVES: The objective of the study was to restore Tc-99m methylene diphosphonate (MDP) bone scan image using blind deconvolution (BD) algorithm so that ribs, vertebrae, and lesions present in them become prominent. MATERIALS AND METHODS: Our study consists of retrospective data in which 356 Tc-99m MDP bone scan images (178 anterior and 178 posterior) were processed using dynamic stochastic resonance algorithm, block-matching 3D filter, and then restored using BD algorithm. Two nuclear medicine (NM) physicians compared restored image with its input image; they especially lookedfor: (a) improvement in lesions detectability, (b) artifacts if any, (c) deterioration in ribs and vertebra, and (d) contrast enhancement in adjacent vertebra and adjacent ribs. They selected one out of two (restored and input) images, which had better quality. The overall image quality was also assessed using the following image quality metrics: brightness, blur, global contrast factor, and contrast per pixel. The Wilcoxon signed-rank test was applied for finding significant difference between the value of image quality metrics of restored image and input image at level of significance alpha = 0.05. RESULTS: According to NM physicians, 80.3% (286 out of 356) of restored images were acceptable, whereas 19.6% (70 out of 356) were unacceptable. Ribs and vertebrae were prominent in 161 out of 178 posterior restored images. Lumbar vertebrae were enhanced and well differentiated from adjacent vertebrae in 125 out of 178 anterior restored images. The value of image quality metrics of restored and input images were found to be significantly different ( P -value < 0.0001). CONCLUSION: Ribs, vertebrae, and lesions present in them become prominent in the most of Tc-99m MDP bone scan images (80.3%) restored using BD algorithm.

Optimizing input parameter of the randomized singular value decomposition algorithm for compressing technetium-99m L,L, ethylenedicysteine renal dynamic study in minimum time preserving clinical information.

OBJECTIVE: The SVDsketch [MATLAB function which implements a randomized singular value decomposition (rSVD) algorithm] uses tolerance (tol) to adaptively determine the rank of the matrix sketch approximation. As the tol gets larger, fewer features of input image matrix are used in the matrix sketch. The objective of this study was to optimize the value of tol for compressing technetium-99m (Tc-99m) L,L, ethylenedicysteine (LLEC) renal dynamic (RD) study in minimum time preserving clinical information. MATERIALS AND METHODS: At different values of tol [0.00012(default), 0.1, 0.01, and 0.05] 50 Tc-99m LLEC RD studies were compressed. Two nuclear medicine (NM) physicians compared compressed images at tol = 0.1 with its input images. The SVD computation time and compression factor were calculated for each study. The image quality metrics: Error, structural similarity index for measuring image quality, brightness, global contrast factor (GCF), contrast per pixel (CPP), and blur were used for objective assessment of image quality. Percentage error in split function estimated from compressed and original images was calculated. Wilcoxon signed-rank test was applied to find statistically significant difference between renal split function, blur, GCF, CPP, and brightness of the compressed image and the original image at . RESULTS: As per NM physicians, compressed images estimated with tol = 0.1 were identical to the original images. Based on image quality metrics, compressed images were significantly less noisy, brighter, and have better contrast compared with its input images. There was insignificant difference in split renal function estimated from compressed RD study at tol = 0.1 and its original study. The SVD computation and percentage compression per study were found to be 0.04725 s and up to 74.53%. CONCLUSION: The optimized value of tol for compressing Tc-99m LLEC RD study preserving clinical information was found to be 0.1, and SVD computation time: 0.04725 s.

Automated Detection of Poor-Quality Scintigraphic Images Using Machine Learning.

Objective In the present study, we have used machine learning algorithm to accomplish the task of automated detection of poor-quality scintigraphic images. We have validated the accuracy of our machine learning algorithm on 99m Tc-methyl diphosphonate ( 99m Tc-MDP) bone scan images. Materials and Methods Ninety-nine patients underwent 99mTC-MDP bone scan acquisition twice at two different acquisition speeds, one at low speed and another at double the speed of the first scan, with patient lying in the same position on the scan table. The low-speed acquisition resulted in good-quality images and the high-speed acquisition resulted in poor-quality images. The principal component analysis (PCA) of all the images was performed and the first 32 principal components (PCs) were retained as feature vectors of the image. These 32 feature vectors of each image were used for the classification of images into poor or good quality using machine learning algorithm (multivariate adaptive regression splines [MARS]). The data were split into two sets, that is, training set and test set in the ratio of 60:40. Hyperparameter tuning of the model was done in which five-fold cross-validation was performed. Receiver operator characteristic (ROC) analysis was used to select the optimal model using the largest value of area under the ROC curve. Sensitivity, specificity, and accuracy for the classification of poor- and good-quality images were taken as metrics for the performance of the algorithm. Result Accuracy, sensitivity, and specificity of the model in classifying poor-quality and good-quality images were 93.22, 93.22, and 93.22%, respectively, for the training dataset and 86.88, 80, and 93.7%, respectively, for the test dataset. Conclusion Machine learning algorithms can be used to classify poor- and good-quality images with good accuracy (86.88%) using 32 PCs as the feature vector and MARS as the classification model.

Restoration of Tc-99m methyl diphosphonate bone scan image using Richardson-Lucy algorithm.

INTRODUCTION: In this study, the optimal input parameters point spread function (PSF) and the number of iterations of the Richardson-Lucy algorithm were experimentally determined to restore Tc-99 m methyl diphosphonate (MDP) whole-body bone scan images. MATERIALS AND METHODS: The experiment was performed on 60 anonymized Tc-99 m MDP whole-body bone scan images. Ten images were used for estimating the optimum value of PSF and the number of iterations to restore scintigraphic images. The remaining 50 images were used for validation of estimated parameters. The image quality of observed and restored images was assessed objectively using blind/referenceless image spatial quality evaluator (BRISQUE), mean brightness (MB), discrete entropy (DE), and edge-based contrast measure (EBCM) image quality metrics. Image quality was subjectively assessed by two nuclear medicine physicians (NMPs) by comparing the restored image quality with observed image quality and assigning a score to each image on the scale of 0-5. RESULTS: Based on BRISQUE, MB, DE, and EBCM scores, the restored images were significantly sharper, less bright, had more detailed information, and had less contrast around edges compared to the input images. The restored images had improved resolution based on visual assessment as well; NMPs assigned an average image quality score of 4.00 to restored images. Maximum resolution enhancement was noticed at PSF (size: 11 pixels, sigma: 1.75 pixels) and the number of iterations = 10. With the increase in the number of iterations, noise also gets amplified along with resolution enhancement and affects the detectability of small lesions; in the case of relatively low noisy input images, the number of iterations = 5 gave better results. CONCLUSION: Tc-99 m MDP bone scan images were restored to improve image quality using the Richardson-Lucy algorithm. The optimum value of the PSF parameter was found to be of size = 11 pixels and sigma = 1.75 pixels.

Near-lossless compression of PET/CT images using singular value decomposition.

AIMS AND OBJECTIVES: The aim of the study is to compare the single matrix approach and slice-by-slice approach for computing singular value decomposition (SVD) to achieve near-lossless compression of PET/CT images. MATERIALS AND METHODS: The parameters used for comparison were SVD computation time, percentage compression and percentage difference between ROI counts on compressed and original images. SVD of 49 F-18-FDG PET/CT studies (33 370 PET/CT images) was computed using both approaches. The smaller singular values contributing insignificant information to the image were truncated, and then, the compressed image was reconstructed. A mask (101 × 101pixels) was used to extract the ROI counts from compressed and original images. Two nuclear medicine physicians compared compressed images with their corresponding original images for loss of clinical details and the presence of generated artifacts. Structural Similarity Index Measure, blur, brightness, contrast per pixel and global contrast factor were used for objective assessment of image quality. Wilcoxon test was applied to find a statistically significant difference between the parameters used for comparison at alpha = 0.05. RESULTS: Nuclear medicine physicians found compressed image identical to the corresponding original image. The values of comparation parameters were significantly larger for the single matrix approach in comparison with the slice-by-slice approach. The maximum percentage error between the compressed image and original image was less than 5%. CONCLUSIONS: Up to 64 % and 44% near-lossless compression of PET and CT images were achieved, respectively using the slice-by-slice approach, and up to 58 and 53% near-lossless compression of PET and CT images were achieved respectively using the single matrix approach.

Restoration of I-131 whole body image using a Wiener filter.

INTRODUCTION: The aim of the study was to restore I-131 whole body image using Wiener filter. MATERIAL AND METHODS: A set of 50 I-131 whole body images acquired using Symbia E dual head gamma camera with high energy general purpose collimator was used. The Gaussian point-spread function (PSF) characterised by the size (3, 5, 7, 9, 11, and 13 pixels) and corresponding standard deviation (0.5, 0.75, 1, 1.5, 1.75, and 2 pixels) and noise-to-signal power ratios (NSR: 0, 0.001, 0.01, 0.1, 0.2, 0.3, 0.4, and 0.5) were used as parameters for Wiener filter. Using the combinations of PSF and NSR, a total of 2450 images (50 × 49 = 2450 images, where 49 images include 1 input and 48 restored images for each input image) were generated and inspected by two nuclear medicine physicians. They selected one best image (the image which had less noise and better contrast between the lesion and background in comparison with the input image). Their results were analyzed. RESULTS: Compared to input image, the metastatic uptake in restored images was very easily perceived. The restored image obtained with PSF (size = 13, sigma = 2) and NSR = 0.3 had better signal-to-noise ratio in comparison to restored image obtained with PSF (size = 11, sigma = 1.75) and NSR = 0.2. CONCLUSION: The restored images with PSF (size = 13, sigma = 2) and NSR = 0.3 were found to have superior image quality in comparison with its input image.

Denoising of iodine-131 images using a median filter.

OBJECTIVE: An iodine-131 (I) image visually appears to be contaminated with impulse noise. The two-dimensional median filter removes noise without sacrificing the edge information. Its performance depends on the shape and size of the mask. In this study, we have compared the performance of a plus-shape and a square-shape median filter for I whole-body images and found the filter with optimum parameter that improves I image quality acceptable to nuclear medicine physicians. MATERIALS AND METHODS: A total of 150 whole-body I images were exported in DICOM format. These images were converted into PNG format and processed with a plus-shape and a square-shape median filter, with each shape mask having sizes of 3, 5, 7, and 9 pixels. The quality of the processed images was assessed by visual assessment by two nuclear medicine physicians and also quantitatively by evaluating metrics: mutual information, mean square error, peak signal-to-noise ratio, and difference entropy. Nuclear medicine physicians assigned a score to each image on the scale 1 (lowest) to 5 (highest) for image quality on the basis of the noise removal, smoothness, and edge information available in the image. Student's t-test was carried out to find the significant difference in the image quality (α=0.05) between the processed images with square-shape and cross-shape mask with the same pixel size. All experiments including statistical analysis were conducted using R installed on a personal computer. RESULTS: Both median filters improved the image quality of I images. The plus-shape median filter was found to show better performance in comparison with the square-shape median filter (P<0.001). The plus-shape median filter with a mask size of 7 pixels was found to be optimum for the processing of whole-body I images. CONCLUSION: The plus-shape median filter with a mask size of 7 pixels can be used to process whole-body I scintigraphic images without loss of clinical information.

Identification of optimal mask size parameter for noise filtering in 99mTc-methylene diphosphonate bone scintigraphy images.

Tc-methylene diphosphonate (Tc-MDP) bone scintigraphy images have limited number of counts per pixel. A noise filtering method based on local statistics of the image produces better results than a linear filter. However, the mask size has a significant effect on image quality. In this study, we have identified the optimal mask size that yields a good smooth bone scan image. Forty four bone scan images were processed using mask sizes 3, 5, 7, 9, 11, 13, and 15 pixels. The input and processed images were reviewed in two steps. In the first step, the images were inspected and the mask sizes that produced images with significant loss of clinical details in comparison with the input image were excluded. In the second step, the image quality of the 40 sets of images (each set had input image, and its corresponding three processed images with 3, 5, and 7-pixel masks) was assessed by two nuclear medicine physicians. They selected one good smooth image from each set of images. The image quality was also assessed quantitatively with a line profile. Fisher's exact test was used to find statistically significant differences in image quality processed with 5 and 7-pixel mask at a 5% cut-off. A statistically significant difference was found between the image quality processed with 5 and 7-pixel mask at P=0.00528. The identified optimal mask size to produce a good smooth image was found to be 7 pixels. The best mask size for the John-Sen Lee filter was found to be 7×7 pixels, which yielded Tc-methylene diphosphonate bone scan images with the highest acceptable smoothness.