Abstract
Objectives
Evaluate a novel algorithm for noise reduction in obese patients using dual-source dual-energy (DE) CT imaging.
Methods
Seventy-nine patients with contrast-enhanced abdominal imaging (55 women; age: 58 ± 14 years; BMI: 39 ± 5 kg/m2, range: 35 – 62 kg/m2) from seven DECT (Siemens SOMATOM Flash or Force) were retrospectively included (01/2019 – 12/2020). Image domain data were reconstructed with the standard clinical algorithm (ADMIRE/SAFIRE 2), and denoised with a comparison (ME-NLM) and a test algorithm (rank-sparse kernel regression). Contrast-to-noise ratio (CNR) was calculated. Four blinded readers evaluated the same original and denoised images (0 (worst) – 100 (best)) in randomized order for perceived image noise, quality, and their comfort making a diagnosis from a table of 80 options. Comparisons between algorithms were performed using paired t-tests and mixed-effects linear modeling.
Results
Average CNR was 5.0 ± 1.9 (original), 31.1 ± 10.3 (comparison; p<0.001), 8.9 ± 2.9 (test; p<0.001). Readers were in good to moderate agreement over perceived image noise (ICC: 0.83), image quality (ICC: 0.71) and diagnostic comfort (ICC: 0.6). Diagnostic accuracy was low across algorithms (accuracy: 66, 63 and 67% (original, comparison, test)). The noise received a mean score of 54, 84, 66 (P<0.05), image quality 59, 61, 65 and the diagnostic comfort 63, 68 and 68, respectively. Quality and comfort scores were not statistically significantly different between algorithms.
Conclusions
The test algorithm produces quantitatively higher image quality than current standard and existing denoising algorithms in obese patients imaged with DECT and readers show a preference for it.
Keywords: Multidetector Computed Tomography, Tomography, X-Ray Computed, Medical Image Processing, Obesity, Image Quality Enhancement
Introduction
Imaging of large body habitus patients is challenging due to high image noise and can lead to delays in patients’ diagnosis and treatment. Many practices do not perform dual-energy (DE) CT on obese patients (e.g., excluding patients weighing over 260 pounds (>118 kg)) due to the additional photon starvation caused by the large body circumference [1].
Dose constraints on DECT scans of obese patients result in high levels of noise, due to photon-starvation effects, negatively impacting diagnostic performance in abdominal radiology [2; 3].
One potential solution to mitigate against noise in larger body habitus populations is by using denoising algorithms. However one concern regarding the implementation of non-vendor based denoising algorithms in clinical practice is the fact that the original projection data is needed for the algorithms to be applied. Since these are very large data files, they do not routinely get stored on the scanners themselves or in PACS and are deleted within approximately a week of acquisition at most institutions [4; 5]. Thus, denoising using projection data, both with vendor-based iterative reconstruction algorithms and custom algorithms, is often limited by data availability. In addition, using iterative reconstruction, there are limitations in spatial resolution for low-contrast structures and low-contrast detectability declines when radiation dose reductions exceed 25% [6]. By contrast, the image domain data from CT scans is usually archived in a PACS system for a relatively long term of several years, so denoising can be applied at any time after image data was acquired and reconstructed. This data can be heterogeneous due to different frequency filters, collimation, slice thicknesses and convolution kernels, but this can be addressed through specialized algorithms.
A denoising method developed at Duke Radiology for the review process is based on rank-sparse kernel regression (RSKR) and applies locally adaptive noise estimation [7]. Other methods apply adaptive non-local means (NLM) filtering by adapting the strength of regularization using a noise variance map generated by propagating noise through the reconstruction model to denoise images. This can lead to a highly smoothed image impression but has been shown to improve image quality in animal and cadaver studies [8; 9].
The purpose of this study was to evaluate a novel algorithm for noise reduction in obese patients for dual-source dual-energy (DE) CT abdominal imaging. We hypothesized that multi-channel denoising methods (RSKR, multi-energy (ME)-NLM) could reduce image noise and improve both objective image quality metrics like contrast-to-noise ratio (CNR) and subjective metrics like reader satisfaction.
Materials and Methods
This retrospective study was Health Insurance Portability and Accountability Act compliant and approved, with informed consent waived by the institutional review board.
Study population
All patients with a BMI of 35 kg/m2 or higher, found during a query of the Duke Clinical Imaging Physics Group (CIPG) Database, with examinations that included a dual-energy acquisition of the abdomen on a dual-source scanner (SOMATOM Flash or Force, Siemens Healthineers, Forchheim, Germany) between January 1st 2019 and December 31st 2020 were consecutively included in the study population (Figure 1). Seventy-nine patients with DECT imaging from seven different scanners (55 women; age: 58 ± 14 years; BMI: 39 ± 5 kg/m2, range: 35 to 62 kg/m2) were included (Table 1). The scan protocols and patient radiation doses varied depending on the scanner type and location within the hospital system (Table 2).
Figure 1:
The patient selection process for inclusion in our study.
Table 1:
Demographic description of patient cohort
| Demographic Parameter | Mean ± SD |
|---|---|
| Age (years) | 58 ± 14 |
| Height (cm) | 165 ± 12 |
| Weight (kg) | 108 ± 21 |
| BMI (kg/m2) | 39 ± 5 |
| Women (number) | 54 |
| Men (number) | 25 |
Table 2:
Shows the distribution of scanners within the hospital, the protocol distribution on each scanner type, and the radiation dose by scanner type.
| Somatom Flash | Somatom Force | |
|---|---|---|
| Number of scanners | 5 | 2 |
| Located in emergency department | 1 | 0 |
| Located in Cancer Center | 2 | 1 |
| Located in non-specialized settings | 2 | 1 |
| Abdomen-Pelvis protocols | 28 | 0 |
| GI protocols | 2 | 1 |
| GU protocols | 6 | 8 |
| Vascular protocols | 15 | 19 |
| Mean CTDIvol (mGy) | 15.2 ± 7.3 | 18.9 ± 6.7* |
P<0.05
DECT acquisition parameters
Dual-source, dual-energy CT of the abdomen was acquired on five SOMATOM Flash CT scanners with tube A set to 100 kVp and tube B at 140 kVp plus tin filtration and on two SOMATOM Force CT scanners with tube A set to 100 kVp and tube B 150 kVp plus tin filtration with a helical pitch of 0.8. Automated exposure control was used: CARE Dose with 190 quality reference mAs and CARE kV with 120 reference kV and 100–140 or 150 kV range depending on scanner type. Tube A had a field of view of 50 cm and tube B of 33 or 35 cm in the Flash and the Force scanners. All patients received 150 ml of iodinated contrast (Isovue 300, Bracco Diagnostics) at an injection rate of 4 ml/sec. Images were the acquired using bolus tracking for portal-venous phase imaging, monitoring was performed in the descending aorta and in multi-phase imaging, triggering for arterial phase images occurred at 150 HU, no further scan delay was added. Portal-venous phase was acquired 80 seconds after the initial trigger.
DECT reconstruction settings
Images were reconstructed at 1.5 mm or 0.6 mm, initially, and displayed for readers using a 1.5 mm slice and 1 mm overlap, the same as in our clinical workflow. Additional reconstruction parameters: Br40 kernel; ADMIRE/SAFIRE 2; 512×512 matrix size with associated in-plane pixel size of 0.67–0.98 mm; collimation of 19.2 or 38.4 mm at isocenter. The spectral post-processing for denoising was performed with the RSKR algorithm (test) as well as with a comparison algorithm from the literature (ME-NLM) [9] using data from tube A and tube B jointly. All images were displayed for readers as blended images with a 50/50 ratio from tube A and tube B.
The hyperparameters for denoising were chosen for RSKR via a preliminary reader preference evaluation, which was based on two radiologists trained at two different institutions in the US and Europe (filtering strength, h0 = 1.2; kernel size: 133 voxels; filter scaling, γ = 0.35; Figure 2). This additional step was taken to ensure that changes in NPS do not affect perceived image texture [10]. For the ME-NLM, hyperparamters were taken from the reference paper (search window, WΩ = 113 voxels; patch size, Wp = 33 voxels; filtering strength, h = 1.2). The same noise map, derived from the local Median Absolute Deviation (MAD) of image gradients, was used for both denoising algorithms.
Figure 2:
Rank-Sparse Kernel Regression (RSKR) is a denoising algorithm designed for multi-energy CT data. (1) Example weighted singular value decomposition of dual-energy data which yields a high-fidelity average image and a lower fidelity difference image. (2–3) Denoising is performed by applying joint bilateral filtration, an edge preserving smoothing operation which matches edges between the average and difference images. Step 2 is alternated with fidelity updates (3), to prevent loss of image details, and further iterated as needed. (4) Denoised results following inversion of the singular value decomposition.
Quantitative image quality analysis
Quantitative metrics were computed and compared between the original, comparison, and test algorithms. Contrast-to-noise ratio (CNR) was calculated for paraspinal skeletal muscle and ventral subcutaneous fat at the same level, chosen because it was present in all images, via the formula: CNR = |attenuationmuscle-attenuationfat|/sqrt(SDmuscle2 + SDfat2). Measurements were made by DPC, with measurements defined in fat and muscle using intensity cut-off values of −200 to −50 HU for fat and −20 to 120 for soft tissue, including all voxels in the specified intensity ranges. The included voxels were applied in identical fashion to all three algorithms.
Modulation transfer function (MTF) measurements to assess spatial resolution were taken at the skin-air boundary, not accounting for changes in in-plane-pixel size. The MTF curve provides a measure of how well the system transfers the contrast ratio of the original object (y-axis) across spatial frequencies (x-axis).
2D noise power spectrum (NPS) measurements were taken to comprehensively assess noise properties. Measurements were made in 322 voxel homogenous regions of interest in subcutaneous fat to assess for objective noise levels. The NPS curve shows the variance in image intensity over spatial frequencies in an image. MTF and NPS measurements were averaged over a subset of 18 patients with matching reconstruction parameters (Flash scanner, 1.5 mm slice thickness with 0.98 mm in-plane-voxel size) using the same measurement regions between algorithms.
Qualitative image quality analysis
Reference standard
To enable randomized and blinded reading, one representative slice from each examination was chosen based on the main diagnosis in the impression of the clinical radiology report. Where exact slice positions were dictated, these were used for slice selection. In cases without a dictated slice position, one radiology fellow (F.R.S., 9 years of experience) selected the slice that showed the main diagnosis best, based on the standard clinical image reconstruction. In cases without pathology, a slice was chosen at the porta hepatis. Each slice was reconstructed with the regular clinical reconstruction parameters, and denoised with the comparison algorithm and the test algorithm. The slice window level was adjusted to produce the most similar visual impression between slices (most used: 300/50 and 650/225) and readers were not able to adjust these during their assessment.
Readers
Slices were presented to four readers not involved in image selection (5, 7, 8 and 9 years of experience) in randomized order in three separate sets that each included all denoising conditions. Each patient was represented once per set and readers completed one set at a time, coming back to the same set whenever needed. They were not allowed to change scores on a previous image set, once it had been completed.
Readers were provided with a reference document for scoring. They were asked to provide scores from 0 to 100 (higher score = better evaluation) for image noise, general image quality and their comfort in making a diagnosis based on the image provided.
In addition, readers were asked to make a diagnosis based on a table with 80 possible diagnoses (Table 3, adapted from Jaffe et al. [11]), which included all diagnoses from the imaging reports. More than one diagnosis was counted as correct, if they were mentioned in the report and visible on the image. Diagnoses made by readers were counted as correct, if at least one of the diagnoses they entered was correct and they were forced to make a diagnosis before they could complete the reads. Overall, there were 948 reads (79 patient cases * 3 denoising conditions * 4 readers).
Table 3:
Table with the diagnoses that were options for the readers to choose from. 0 was considered as “no pathology”.
| Organ system | Diagnosis A | Diagnosis B | Diagnosis C | Diagnosis D | Diagnosis E |
|---|---|---|---|---|---|
| Lung bases | Atelectasis (1) | Consolidation (2) | Effusion (3) | Nodule (4) | Fibrosis (5) |
| Cardiac | Cardiomegaly (6) | Effusion (7) | Coronary Calcium (8) | Pacemaker (9) | Enlarged (10) |
| Liver | Mass (11) | Fat(12) | Enlarged (13) | Cirrhosis (14) | Ascites (15) |
| Gallbladder | Absent (16) | Stone/Sludge (17) | Wall thickening (18) | Fluid (19) | Mass (20) |
| Biliary tree | Duct dilatation (21) | Stone (22) | Mass (23) | Pneumobilia (24) | / |
| Pancreas | Mass (25) | Stranding (26) | Duct dilatation (27) | Calcification (28) | Fluid (29) |
| Spleen | Mass (30) | Enlarged (31) | Calcification (32) | Infarct (33) | Splenule (34) |
| Adrenals | Mass (35) | Calcification (36) | Hyperplasia (37) | / | / |
| Kidneys | Mass (38) | Stone (39) | Hydronephrosis (40) | Pyelonephritis (41) | Infarct (42) |
| Ureters | Mass (43) | Stone (44) | Dilatation (45) | / | / |
| Stomach | Mass (46) | Wall thickening (47) | Ulcer (48) | Distension (49) | / |
| Small Bowel | Mass (50) | Wall thickening (51) | Ileus (52) | Dilatation (53) | Obstruction (54) |
| Colon | Mass (55) | Wall thickening (56) | Diverticulosis (57) | Diverticulitis (58) | Appendicitis (59) |
| Lymph nodes | Increase size (60) | Increase number (61) | Calcification (62) | / | / |
| Arteries | Aneurysm (63) | Thrombosis (64) | Calcification (65) | Variants (66) | Stent(67) |
| Veins | Varices (68) | Thrombosis (69) | Calcification (70) | Variants (71) | Stent(72) |
| Bones | Mass (73) | Infection (74) | Fracture (75) | / | / |
| Others | Surgical changes (76) | Ovarian mass (77) | Fat stranding (78) | Atrophy (79) | Hernia (80) |
Statistical analysis
Descriptive statistics were used to characterize the patient population. Image scores were compared using paired two-tailed t-tests. Reader accuracy was compared using one-way ANOVA with Bonferroni correction. Reader scores were compared per denoising algorithm and the assumptions for linear models were confirmed. To avoid biasing the results with clustering by readers or patients, scores were compared between algorithms using mixed-effects linear regression with denoising algorithm treated as a fixed effect and patient and reader as random effects. Post-hoc testing was performed using the Tukey-test. Inter-reader agreement was determined using intraclass correlation coefficients (ICC) and Spearman’s rank correlation test. The accuracy of diagnosis was determined based on the reference standard. The level of statistical significance was assumed at alpha = 0.05. Statistical analysis was performed using RStudio Version 1.3.1056.
Results
Quantitative image quality analysis
Average CNR values were 5.0 ± 1.9 (original), 31.1 ± 10.3 (comparison; p<0.001), 8.9 ± 2.9 (test; p<0.001). These quantitative measures are correlated with reader perceptions of noise detailed in the next section. The denoising algorithms did not aversely affect spatial resolution measurements at the skin-air boundary (MTF at 50%: ~3.0 cm−1; MTF at 10%: ~5.2 cm−1 for all algorithms). Both denoising algorithms significantly reduced noise power relative to the original reconstructions at high spatial frequencies and performed well at the creation of iodine maps and virtual non-contrast images (an example is shown in Figure 3). The ME-NLM algorithm removed more noise at mid spatial frequencies than RSKR (Supplemental Figure 1).
Figure 3:
Dual-energy post-processing into iodine and virtual non-contrast (VNC) images for a 55-year-old female patient with a pancreatic tail lesion. Insets highlight improved iodine map image quality following denoising with the test algorithm and the location of the lesion (white arrows).
Qualitative image quality analysis
The perceived image noise received a mean score of 54 ± 21 (original), 84 ± 10 (comparison), 66 ± 17 (test). The difference in perceived image noise was significant between algorithms with the comparison algorithm receiving an average of 29.6 higher scores than the original images (P=0.045) and the test algorithm receiving an average of 11.8 higher scores than the original images (P=0.014; Table 4). There was no statistically significant difference between the scores for the comparison algorithm and the test algorithm (P=0.102). Readers were overall in good agreement over perceived image noise (ICC: 0.83). This finding correlates with the quantitative measures of reduced image noise for both of the denoising algorithms, the perception of the readers agreeing with the NPS measurements.
Table 4:
Mean scores for perceived image noise, image quality and reader confidence in making a diagnosis with 95% Confidence Intervals; bold* statistically significant improvement over standard reconstruction.
| Original | Comparison | Test | |
|---|---|---|---|
| Image noise | 54.0 [21.0–87.0] |
83.6* [68.4–98.7] |
65.8* [34.9–96.7] |
| Image Quality | 59.4 [33.7–85.1] |
61.1 [23.9–98.3] |
64.5 [34.8–94.3] |
| Diagnostic Comfort | 63.4 [38.5–88.4] |
67.5 [34.5–100.5] |
67.8 [37.0–98.6] |
The perceived image quality received a mean score of 59 ± 16 (original), 61 ± 24 (comparison), 65 ± 19 (test). The difference in perceived overall image quality was not statistically significant for the comparison algorithm (average of 1.7 higher, P=0.935) or the test algorithm (average of 5.1 higher, P=0.293). There was no statistically significant difference between the scores for the comparison algorithm and the test algorithm (P=0.480). Readers were in moderate agreement over image quality (ICC: 0.72).
The diagnostic comfort received a mean score of 64 ± 16 (original), 68 ± 21 (comparison) and 68 ± 20 (test). The difference in diagnostic comfort was not statistically significant for the comparison algorithm (average of 4.1 higher, P=0.489) or the test algorithm (average of 4.3 higher, P=0.351). There was no statistically significant difference in diagnostic comfort between the scores for the comparison algorithm and the test algorithm (P=0.986). Readers were in moderate agreement over diagnostic comfort (ICC: 0.66).
Diagnostic accuracy, where readers gave at least one of the diagnoses mentioned in the report, was relatively low across algorithms (accuracy: 66, 63 and 67% for original, comparison, and test; Figure 4). These differences were not statistically significant when comparing accuracy between original and comparison algorithm (P=0.527), original and test algorithm (P=0.741) or the two different denoising algorithms (P=0.202).
Figure 4:
Demonstrates two challenging cases. Case 1 (top row), in which only two of the readers made the diagnosis of left adrenal lesion in a 43 year-old female patient with a BMI of 41.3 kg/m2 on the original reconstruction with standard parameters. Case 2 (bottom row), in which only two of the readers made the diagnosis of thickened gall bladder wall in a 37 year-old female patient with a BMI of 36.2 kg/m2 on the original reconstruction with standard parameters.
There were differences between individual readers in preference for algorithm and the diagnostic accuracy depending on algorithm and reader (Figure 5, Supplemental Figures 2 and 3). Two readers gave the highest diagnostic comfort score to the RSKR, but one showed better diagnostic accuracy on the original images, while one other reader gave the highest diagnostic comfort score to the original images but had the highest diagnostic accuracy with our test algorithm (Table 5, Figure 6).
Figure 5:
The differences in image noise perception by reader and denoising algorithm. The scores for original images are shown in pale blue (left), for the comparison algorithm in light blue (middle) and for the test algorithm in dark blue (right).
Table 5:
Shows the individual scores given by readers for image noise, quality and diagnostic comfort. Highlighted are the best scores per category for each reader.
| Reader 1 | Reader 2 | Reader 3 | Reader 4 | |
|---|---|---|---|---|
| Image noise Original | 35.6 ± 13.2 | 73.0 ± 12.2 | 71.1 ± 10.1 | 36.2 ± 18.2 |
| Image noise Comparison | 77.5 ± 17.7 | 97.3 ± 5.9 | 83.4 ± 8.1 | 76.1 ± 13.9 |
| Image noise Test | 49.9 ± 16.3 | 87.5 ± 12.4 | 77.4 ± 9.3 | 48.5 ± 19.5 |
| Image quality Original | 40.5 ± 14.4 | 76.9 ± 13.1 | 68.5 ± 9.9 | 51.7 ± 20.4 |
| Image quality Comparison | 39.2 ± 15.8 | 92.8 ± 10.0 | 64.0 ± 7.7 | 48.5 ± 19.7 |
| Image quality Test | 47.5 ± 18.6 | 89.0 ± 11.2 | 69.8 ± 9.3 | 52.0 ± 21.2 |
| Diagnostic comfort Original | 49.3 ± 25.7 | 79.6 ± 20.9 | 75.1 ± 16.2 | 49.7 ± 29.8 |
| Diagnostic comfort Comparison | 52.0 ± 25.5 | 92.7 ± 12.0 | 77.1 ± 14.7 | 48.3 ± 31.5 |
| Diagnostic comfort Test | 56.2 ± 26.7 | 89.1 ± 14.5 | 79.0 ± 15.2 | 47.0 ± 33.6 |
Figure 6:
Demonstrates the inter-reader differences in diagnostic accuracy based on whether and which form of denoising was applied to the images. Readers 1 and 3 preferred the images generated by the test algorithm but only Reader 1 also had highest accuracy using these images, while Readers 2 and 4 preferred the overall image quality from the comparison algorithm but had highest accuracy with the test algorithm.
Discussion
Our study showed that diagnostic accuracy was overall low for obese patients and only slightly higher when using our denoising algorithm. The average scores for image noise, image quality and diagnostic comfort were higher than for the original iterative reconstructions without denoising, and the highest image quality scores were given to the RSKR algorithm.
This is similar to results reported for commercially available denoising algorithms applied to single-energy, low-dose CT, where image quality was rated as poor with FBP but moderate when using the denoising algorithm [12].
The comparison algorithm used received the best noise scores, but lower average subjective image quality scores than the test algorithm (61 vs. 65). This may have been due to the fact that the algorithm smoothes out the image texture to a point where radiologists are uncomfortable with the “plastic” appearance of abdominal structures, which has been described for high iterative-reconstruction strengths [13; 14] and may in part be due to changes in NPS [10].
There was some disagreement between which of the denoised results the radiologists preferred and how accurate they were in making their diagnoses. Two readers preferred the RSKR but one was less accurate in their diagnoses with RSKR than with the original reconstruction. One other reader preferred the ME-NLM results but was the most accurate with RSKR. While this is an unexpected effect and we did not record time spent interpreting each image, it might be due to readers spending more time scrutinizing the images they preferred less and thus reaching the correct diagnosis more often. The effect of time spent looking at images and diagnostic accuracy has been shown for mammography data [15], but should be explored further in future studies.
In addition, though there was a general preference for denoised images of any kind, both readers with the highest comfort scores assigned to RSKR were trained at our institution, while the readers who preferred ME-NLM received their initial training at a different institution. There might be an effect of diagnostic comfort being dictated by an image that is closer to what the training institution was using (e.g. iterative reconstruction vs. filtered back projection as default), since protocols vary widely between institutions nationally and internationally [16; 17].
Recently, radiology reader fatigue has come to be a topic of some research interest and it might be reasonable to assume that looking at images of higher quality could lead to less reader fatigue over the course of a work day even if it does not improve accuracy on a per-scan basis, which should be evaluated in future studies [18].
Though overall diagnostic accuracy was not statistically significantly improved, using denoising algorithms might be a way of reducing radiation dose in obese patients, which should be explored in prospectively designed future studies. Since this denoising technique was added to the standard iterative reconstruction already used in clinical practice it would also be important to integrate these techniques into the background workflows performed at the scanner or in PACS so using them will not lead to increased workloads for technologists or radiologists. This would also increase the availability of the denoising techniques to obese patients in all of the United States and potentially world-wide, instead of it being used only at highly-specialized academic centers.
Limitations
In addition to the retrospective nature, several limitations of our study deserve attention. First our study was limited by the fact that it was a single center study with a relatively small sample size and with variance in the scanner types and acquisition protocols due to the challenges of acquiring CT data from obese patients. Second, additional improvement in diagnostic quality might be achieved by providing readers with the full dataset and clinical history, which should be explored in follow-up investigations, since the task of making a diagnosis based on a single CT slice is particularly difficult and may have contributed to the overall low diagnostic accuracy in our study. Third, the retrospective nature of data acquisition does not provide the opportunity to draw direct conclusions about clinical management; this should be further investigated in a prospective study designed to test clinical performance of denoising algorithms in obese patients. The reader study was limited to a comparison between the standard iterative reconstructions (ADMIRE/SAFIRE 2) and a non-local-means based algorithm (ME-NLM). In the future, it might be feasible to broaden the comparisons to deep learning algorithms and potentially evaluate denoising for virtual monoenergetic imaging.
For future directions, since our algorithm works in the image domain and is not specific to an acquisition method, we believe it would be applicable for the emerging technology of photon-counting CT data, which may also profit from denoising algorithms. In a recent study, the RSKR denoising algorithm was applied during iterative reconstruction to compare preclinical micro-CT data using energy-integrating detectors and a photon-counting detector with four energy thresholds to image iodine and gadolinium based contrast agents in sarcoma tumors induced on the flanks of mice [19]. The combination of inherently higher contrast resolution in photon-counting CT and denoising algorithms could further improve imaging of obese patients.
In conclusion, the new algorithm produces quantitatively higher image quality data than current standard and existing denoising algorithms in obese patients imaged with DECT. Readers show a preference for it and their average diagnostic accuracy has a tendency for improvement.
Supplementary Material
Supplemental Figure 3: Visualizes the differences in diagnostic confidence by reader and denoising algorithm. The scores for original images are shown in pale blue (left), for the comparison algorithm in light blue (middle) and for the test algorithm in dark blue (right).
Supplemental Figure 2: Visualizes the differences in image quality perception by reader and denoising algorithm. The scores for original images are shown in pale blue (left), for the comparison algorithm in light blue (middle) and for the test algorithm in dark blue (right).
Supplemental Figure 1: Modulation transfer and radial NPS measurements averaged over 18 patients.
Key points.
Improving image quality in DECT imaging of obese patients is important for accurate and confident clinical reads, which may be aided by novel denoising algorithms using image domain data.
Accurate diagnosis on CT imaging of obese patients is especially challenging and denoising algorithms can increase quantitative and qualitative image quality.
Image domain algorithms can generalize well and can be implemented at other institutions.
Acknowledgements
This project was made possible by a collaborative research agreement with Siemens Healthineers (Erlangen, Germany). We would like to thank the Duke Clinical Imaging Physics Group under Dr. Samei for their support in identifying patients eligible for our study.
List prior or ongoing funding for this project. List N/A if not applicable
This study has received funding by the NIH National Cancer Institute (U24 CA220245, R01 CA196667 and RF1AG070149-01).
Abbreviations
- BMI
Body-Mass-Index
- CIPG
Clinical Imaging Physics Group
- CNR
Contrast-to-Noise Ratio
- DECT
Dual-Energy CT
- ICC
Intra Class Correlation
- kVp
kilovolt peak
- MAD
Median Absolute Deviation
- ME-NLM
Multi-Energy Non-Local Means
- ml
milliliter
- MTF
Modulation Transfer Function
- NPS
Noise Power Spectrum
- PACS
Picture Archiving and Communication System
- ROI
Region Of Interest
- RSKR
Rank Sparse Kernel Regression
- SD
Standard Deviation
- sec
second
References
- 1.Megibow AJ, Sahani D (2012) Best Practice: Implementation and Use of Abdominal Dual-Energy CT in Routine Patient Care. American Journal of Roentgenology 199:S71–S77 [DOI] [PubMed] [Google Scholar]
- 2.Modica MJ, Kanal KM, Gunn ML (2011) The Obese Emergency Patient: Imaging Challenges and Solutions. Radiographics 31:811–823 [DOI] [PubMed] [Google Scholar]
- 3.Uppot RN (2018) Technical challenges of imaging & image-guided interventions in obese patients. The British Journal of Radiology 91:20170931. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Rubin GD (2000) Data explosion: the challenge of multidetector-row CT. European Journal of Radiology 36:74–80 [DOI] [PubMed] [Google Scholar]
- 5.Langer SG (2011) Challenges for Data Storage in Medical Imaging Research. Journal of Digital Imaging 24:203–207 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Mileto A, Guimaraes LS, McCollough CH, Fletcher JG, Yu L (2019) State of the Art in Abdominal CT: The Limits of Iterative Reconstruction Algorithms. Radiology 293:491–503 [DOI] [PubMed] [Google Scholar]
- 7.Clark DP, Badea CT (2017) Hybrid spectral CT reconstruction. PLoS One 12:e0180324. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Li Z, Yu L, Trzasko JD et al. (2014) Adaptive nonlocal means filtering based on local noise level for CT denoising. Med Phys 41:011908. [DOI] [PubMed] [Google Scholar]
- 9.Li Z, Leng S, Yu L, Manduca A, McCollough CH (2017) An effective noise reduction method for multi-energy CT images that exploit spatio-spectral features. Medical Physics 44:1610–1623 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Greffier J, Frandon J, Larbi A, Beregi J, Pereira F (2020) CT iterative reconstruction algorithms: a task-based image quality assessment. European Radiology 30:487–500 [DOI] [PubMed] [Google Scholar]
- 11.Jaffe TA, Martin LC, Miller CM et al. (2007) Abdominal pain: coronal reformations from isotropic voxels with 16-section CT--reader lesion detection and interpretation time. Radiology 242:175–181 [DOI] [PubMed] [Google Scholar]
- 12.Steuwe A, Weber M, Bethge OT et al. (2021) Influence of a novel deep-learning based reconstruction software on the objective and subjective image quality in low-dose abdominal computed tomography. The British Journal of Radiology 94:20200677. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Mohammadinejad P, Mileto A, Yu L et al. (2021) CT Noise-Reduction Methods for Lower-Dose Scanning: Strengths and Weaknesses of Iterative Reconstruction Algorithms and New Techniques. Radiographics 41:1493–1508 [DOI] [PubMed] [Google Scholar]
- 14.Geyer LL, Schoepf UJ, Meinel FG et al. (2015) State of the art: iterative CT reconstruction techniques. Radiology 276:339–357 [DOI] [PubMed] [Google Scholar]
- 15.Carney PA, Bogart TA, Geller BM et al. (2012) Association between time spent interpreting, level of confidence, and accuracy of screening mammography. AJR American journal of roentgenology 198:970–978 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Willatt JMG, Mason AC (2006) Comparison of radiology residency programs in ten countries. European Radiology 16:437–444 [DOI] [PubMed] [Google Scholar]
- 17.Homayounieh F, Holmberg O, Umairi RA et al. (2021) Variations in CT Utilization, Protocols, and Radiation Doses in COVID-19 Pneumonia: Results from 28 Countries in the IAEA Study. Radiology 298:E141–E151 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Taylor-Phillips S, Stinton C (2019) Fatigue in radiology: a fertile area for future research. The British Journal of Radiology 92:20190043. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Badea CT, Clark DP, Holbrook M, Srivastava M, Mowery Y, Ghaghada KB (2019) Functional imaging of tumor vasculature using iodine and gadolinium-based nanoparticle contrast agents: a comparison of spectral micro-CT using energy integrating and photon counting detectors. Phys Med Biol 64:065007. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Supplemental Figure 3: Visualizes the differences in diagnostic confidence by reader and denoising algorithm. The scores for original images are shown in pale blue (left), for the comparison algorithm in light blue (middle) and for the test algorithm in dark blue (right).
Supplemental Figure 2: Visualizes the differences in image quality perception by reader and denoising algorithm. The scores for original images are shown in pale blue (left), for the comparison algorithm in light blue (middle) and for the test algorithm in dark blue (right).
Supplemental Figure 1: Modulation transfer and radial NPS measurements averaged over 18 patients.