The efficacy of 2D,non-linear noise reduction filtering in cardiac imaging: a pilot study

Abstract

Background

Computed tomography (CT) is becoming increasingly popular as a non-invasive method for visualizing the coronary arteries but patient radiation doses are still an issue. Postprocessing filters such as 2D adaptive non-linear filters might help to reduce the dose without loss of image quality.

Purpose

To investigate whether the use of a 2D, non-linear adaptive noise reduction filter can improve image quality in cardiac computed tomography angiography (CCTA).

Material and Methods

CCTA examinations were performed in 36 clinical patients on a dual source CT using two patient dose levels: maximum dose during diastole and reduced dose (20% of maximum dose) during systole. One full-dose and one reduced-dose image were selected from each of the examinations. The reduced-dose image was duplicated and one copy postprocessed using a 2D non-linear adaptive noise reduction filter, resulting in three images per patient. Image quality was assessed using visual grading with three criteria from the European guidelines for assessment of image quality and two additional criteria regarding the left main artery and the overall image quality. Also, the HU value and its standard deviation were measured in the ascending and descending aorta. Data were analyzed using Visual Grading Regression and paired t-test.

Result

For all five criteria, there was a significant (P < 0.01 or better) improvement in perceived image quality when comparing postprocessed low-dose images with low-dose images without noise reduction. Comparing full dose images with postprocessed low-dose images resulted in a considerably larger, significant (P < 0.001) difference. Also, there was a significant reduction of the standard deviation of the HU values in the ascending and descending aorta when comparing postprocessed low-dose images with low-dose images without postprocessing.

Conclusion

Even with an 80% dose reduction, there was a significant improvement in the perceived image quality when using a 2D noise-reduction filter, though not approaching the quality of full-dose images. This indicates that cardiac CT examinations could benefit from noise-reducing postprocessing with 2D non-linear adaptive filters.

Keywords

Cardiac CT angiography heart adults image manipulation

As faster and faster multidetector computed tomography (MDCT) scanners become available, CT is also becoming increasingly popular as a non-invasive method for visualizing the coronary arteries. However, patient radiation doses are still a concern. In one review of several studies, coronary computed tomography angiography (CCTA) gave an average effective dose of approximately 16 mSv (5–32 mSv), compared to approximately 7 mSv (2–15.8) for conventional diagnostic angiography (1). Even so, CCTA has a number of advantages such as its non-invasive nature, and the possibility to extract additional anatomical and physiological information.

Image acquisition is usually ECG-modulated, with the highest dose used during diastole, and depending on the equipment used, radiation is sequential or continuous. The use of continuous radiation offers an opportunity to evaluate ventricular movement throughout the whole cardiac cycle, but it also increases the radiation dose.

Today, most modern scanners have effective dose-reducing features such as automatic exposure control and in some cases also ECG-gating (2, 3), but there are also several types of postprocessing filters available, such as 2D adaptive non-linear filters (4, 5), that can help to reduce the dose further without loss of image quality. In essence, adaptive non-linear filters work by estimating the local orientation in each pixel and letting this determine what degree of smoothing is applied in that pixel (Fig. 1). Here, the orientation estimation and filtering is carried out in 2D. Thus, the filtering process varies for each separate pixel and is not homogenously applied to the whole image (6). This process is carried out on a separate workstation after reconstruction on the scanner (Fig. 2). The efficacy of 2D adaptive non-linear filters has previously been proved for a number of organs and lesions (7–10), but not for the coronary arteries.

Fig. 1
Principle of adaptive filtering (after [6]). The neighborhood of each pixel is described by computing features such as local orientation, local variance or local phase, and several versions of smoothed or sharpened images are constructed. How these versions are combined in each pixel is determined by the local features and by parameter settings that can be chosen by a competent user in a tuning phase

Fig. 2
Typical image flow. Postprocessing is separate from initial image acquisition and reconstruction

Receiver-operating characteristics (ROC) analysis is generally considered to be the state-of-the art in image quality studies (11). However, although this method can actually measure the ability to detect a pathologic lesion, it also requires a large number of images with reliable reference diagnoses. In contrast, visual grading is a simpler yet powerful method, which is well suited for clinical image quality evaluation. It is based on the assumption that the visibility of important anatomical structures correlates with the ability to detect pathologic lesions. This assumption has also been shown to be valid in a few specific cases where the results of visual grading studies have been compared to those of corresponding ROC studies (12–14).

The aim of this work was to use visual grading in order to investigate whether the application of a 2D, non-linear adaptive noise reduction filter can improve image quality in cardiac computed tomography angiography (CCTA).

Material and Methods

Patients and imaging

All images in this study were anonymized. All were parts of regular exams and warranted no extra radiation. Thus, according to current Swedish legislation, there was no need for an ethical approval.

A total of 47 consecutive CCTA exams on patients referred for chest pains were selected. CCTA was performed using a 64-slice dual source CT (Somatom Definition, Siemens™, Forcheim, Germany) with ECG-modulated tube current and retrospective ECG-gating. Scan protocols were selected according to clinical practice, depending on body mass (Ref. 380 mAs, 100 kV and pitch 0.29 for patients <90 kg. Ref. 310 mAs, 120 kV and pitch 0.25 for patients ≥90 kg; rotation time 0.33 s, collimation 0.6 mm).

During the diastolic phase the maximum (full) radiation dose was used to ensure optimal image quality. During the systolic phase, the dose was reduced to approximately 20% of the maximum dose. The resulting images were presented in 5% steps ranging from 35–75% of the R-R interval. Usually the 35% phase was captured with the reduced, 20% dose, while the 40–75% phases were captured with full dose. However, nine exams were excluded from the study as the 35% phase was captured with the maximum dose, leaving no reduced-dose images for comparison. A further two exams were excluded as a completely different study protocol was used in these cases, leaving a total of 36 exams to be included in the study (17 male patients, 19 female patients, age 31–84 years).

Due to inconsistencies in the protocol selection, the 120 kV protocol was used in six patients weighing less than 90 kg. Thus, a total of 24 patients were examined with the 120 kV protocol (mean 87.8 kg, range 65–108 kg) and 12 with the 100 kV protocol (mean 68.5 kg, range 51–87 kg). The average volume computed tomography dose index (CTDI_vol) was 29.2 (range 13.8–68.7) mGy, with an average dose length product (DLP) of 487 (range 238–1102) mGy×cm.

From each exam, two single 0.75 mm images covering parts of the left main coronary artery were selected, one from the reduced dose, 35% (of the R-R-interval) stack and one from the full dose, 70% stack. The low-dose image was duplicated and one of the copies postprocessed using a separate, hardware based 2D, non-linear adaptive noise reduction filter (SharpView™ CT; Sharpview AB, Linköping, Sweden), resulting in three images per patient (Fig. 3a–c) and a total of 108 images. The filter parameters were set by Sharpview AB prior to the study and were optimised for use with the low-dose cardiac CT images included in this study. The same filter settings were used for all patients, regardless of protocol or body weight, and no adjustments could be made by the observers.

Fig. 3
(a) Transaxial CT image of the section of interest; 100% dose and no postprocessing. (b) Transaxial CT image of the section of interest; 20% dose and no postprocessing. (c) Transaxial CT image of the section of interest; 20% dose and postprocessing

Image analysis

Subsequently, the images were transferred as DICOM-files to a visual grading software package (ViewDEX 1.0; University of Gothenburg, Sweden) for further assessment (15). The quality of each image was independently assessed by nine radiologists, resulting in 972 observations. Appropriate instructions were issued beforehand but apart from that, none had any prior knowledge of the study and its aims. All patient and image data were hidden to the observers. For training purposes, each radiologist assessed an additional 21 images prior to the study material itself.

All image assessment was made on regular, DICOM-calibrated PACS workstations. The window settings were freely adjustable by the observers.

The perceived image quality was graded using a five-point scale for five different image quality criteria, resulting in 4860 parameter ratings.

Three of the image quality criteria used consisted of applicable parts of the European Guidelines for assessment of image quality (16). Another two criteria were added, one regarding the visualization of the left main coronary artery and one regarding the overall image quality.

Thus, the criteria used were as follows:
Visually sharp reproduction of the thoracic aorta;

Visually sharp reproduction of the wall of the thoracic aorta;

Visually sharp reproduction of the heart;

Visually sharp reproduction of the left main coronary artery (LMA);

The image noise in relevant regions is sufficiently low for diagnosis.
The scale used was as follows:
Criterion is fulfilled;

Criterion is probably fulfilled;

Indecisive;

Criterion is probably not fulfilled;

Criterion is not fulfilled.
In order to obtain an objective marker of image noise, the Hounsfield Unit (HU) value and its standard deviation were measured in all images in the ascending aorta (n = 36) and, where available, in the descending aorta (n = 33). The region of interest (ROI) used was drawn in one of the low-dose images and then copied onto the other two images in order to ensure as identical a measurement as possible. The ROI was circular and made to cover approximately two-thirds of the lumen. The average ROI area was 3.8 cm² (range 2.1–6.7 cm²) in the ascending aorta, and 1.3 cm² (range 0.3–2.2 cm²) in the descending aorta.

Statistical analysis

The results of the visual grading were analysed using Visual Grading Regression (VGR) (17), i.e. ordinal logistic regression comparing the three types of images while controlling for patient and observer identity, using JMP 7.0.1 (SAS Inc., Cary, NC, USA).

Analysis of the HU values was made using the Student's paired t-test (SPSS Statistics 17.0, SPSS, Chicago, IL, USA).

Results

The results of the visual grading are summarized for each criterion in Figs. 4–8. For all five criteria, there was a significant improvement in perceived image quality when comparing postprocessed low-dose images with low-dose images without postprocessing (P < 0.001 except criterion 1 where P < 0.01). Comparing full dose images with postprocessed low-dose images resulted in a considerably larger, significant (P < 0.001) difference for all image criteria.

Fig. 4
Criterion 1: Visually sharp reproduction of the thoracic aorta. Perceived image quality graded from 1 (criterion fulfilled) to 5 (criterion not fulfilled)

Fig. 5
Criterion 2: Visually sharp reproduction of the wall of the thoracic aorta. Perceived image quality graded from 1 (criterion fulfilled) to 5 (criterion not fulfilled)

Fig. 6
Criterion 3: Visually sharp reproduction of the heart. Perceived image quality graded from 1 (criterion fulfilled) to 5 (criterion not fulfilled)

Fig. 7
Criterion 4: Visually sharp reproduction of the left main coronary artery (LMA). Perceived image quality graded from 1 (criterion fulfilled) to 5 (criterion not fulfilled)

Fig. 8
Criterion 5: The image noise in relevant regions is sufficiently low for diagnosis. Perceived image quality graded from 1 (criterion fulfilled) to 5 (criterion not fulfilled)

Applying the 2D noise reduction filter to the low-dose images reduced the average standard deviation of the HU value from 70.5 to 59.5 in the ascending aorta (P < 0.001), and from 67.2 to 55.8 in the descending aorta (P < 0.001). The average standard deviation in the full-dose images was 35.2 in the ascending aorta and 34.6 in the descending aorta. There was also an increase in the average HU value of 1.3 in the ascending aorta (P < 0.001), of 1.5 in the descending aorta (P < 0.001) when applying the noise reduction filter to the low-dose images.

Discussion

In later years, CCTA has become a realistic alternative to conventional coronary angiography in selected patient groups (18, 19). Due to the rapid advances made in CT technology, it is now possible to perform a scan of the coronary arteries using a dose of less than 1 mSv, if the circumstances are optimal (20). This is certainly a low dose but it is not negligible, and when circumstances are less than optimal, the dose is still considerably higher. More importantly, it will probably take a number of years before these advanced machines are commonly spread throughout the world. Meanwhile, a host of older 16- and 64-slice CT-scanners will continue to be used for cardiac imaging and with these machines the dose will inevitably be higher, even though most reasonably, modern scanners are equipped with a number of dose-reducing features.

In recent years, promising results concerning dose reduction have been achieved by the use of iterative reconstruction algorithms (IRA) which, during the image reconstruction process, utilize repeated comparisons between estimated and measured image data in order to reduce image noise (21). The method itself is not new, having been around since the very early days of computed tomography, and even longer in the mathematical field, but new algorithms and increased computer power are now beginning to make it clinically useful (22). In contrast to IRA, and also to other dose-reducing measures such as automatic dose control and ECG-gating, the 2D adaptive non-linear filters discussed here are applied on already existing, reconstructed images. Thus, a filter of this kind can be used in addition to other dose-reducing measures.

In this pilot study, we have investigated the effect of a 2D non-linear adaptive noise reduction filter on images with a rather large dose reduction of 80%. The results show that even with this substantial dose reduction, there was significant improvement of the perceived image quality with the application of a 2D non-linear adaptive noise reduction filter. Still, the improvement was small compared to the superior image quality of the full-dose images. Thus the filtering was not able to compensate for the large dose reduction.

Also, there was a significant reduction of image noise expressed as the average standard deviation of the HU value in the ascending and descending aorta. However, even in regions where the HU values are expected to remain constant, such as in large vessels where the contrast medium concentration can be assumed not to vary spatially, the standard deviation of the HU value is not a complete descriptor of noise since it does not include any information of the spatial noise correlations as expressed in terms of noise power spectrum.

The main reason for studying a dose reduction of 80% was that these images were part of regular examinations and readily available without adding any extra dose to the patient. In future studies, it would probably be more interesting to study the effect of similar filters on images with a smaller dose reduction, where it is less obvious whether the postprocessing might compensate fully for the increased image noise. The level of dose reduction that one can expect to achieve by applying filtering can, in fact, be estimated from the results of the VGR analysis in the present study.

An unforeseen consequence of the application of the filter was the slight but significant increase of the average HU values in the aorta. This might be related to the selection of the filtering parameter settings but some kind of error in the noise reduction algorithm cannot be excluded. If this introduces a problem in clinical image evaluation remains to be seen.

There are some limitations to this study. It does not take into consideration to which extent the capability to detect pathology is improved by the utilisation of the filter. However, as previously discussed, this is not possible with our study design. A basic assumption in visual grading studies is that the visibility of anatomic structures also correlates with the detectability of pathologic lesions. To assess the detectability of pathologic lesions in a more direct way, the employment of ROC analysis might be considered, although this would require a rather more complicated set-up with a more carefully selected image material than is present in this study as well as access to undisputable reference diagnoses.

Naturally, there are a number of other weaknesses to this study, the main objection being the use of single images only from multiple image CT exams. This was due to limitations in the viewing software package which at the time was unable to handle image stacks.

Ideally, some other structure than the aorta should have been used for the measurement of HU and their standard deviation value but a small field of view in combination with single slices meant it was difficult to find any homogenous anatomical structures that were consistently depicted in all images. As it is now, the low dose and full dose images were taken from different phases of the cardiac cycle and even though the difference in time is small, there is no way of telling whether the amount of contrast in the two images is equal. However, the effect of the filter is clear as the two low-dose images were identical in all other respects.

A further stratification of the patient material according to BMI/body weight, tube voltage and possibly gender might have been appropriate. An attempt to do so showed that, except for criterion 1, the results seemed to hold when stratifying the patient material according to bodyweight and BMI (higher or lower than the median value). However, the clinical nature of the examinations included also meant that a certain inconsistency was present in the choice of exam protocol. This made it difficult to define homogenous groups which in the end, together with the limited number of observations, made us decide against further stratification. Even though we thus lack information on results in several subpopulations, we feel that for this pilot study, it may suffice to know that the results are valid for the group as a whole.

One could also argue that it would have been appropriate to evaluate the effect of the filter on the full-dose images as well. However, as the number of images to be included was somewhat limited by the time needed to evaluate them, we decided to focus mainly on the potential enhancement of the low-dose images. Also, the full-dose images would require a different filter setting, making direct comparison between the groups difficult.

In conclusion, even with an 80% dose reduction, there was a significant improvement in the perceived image quality when using a 2D noise-reduction filter, though not approaching the quality of full-dose images. This indicates that cardiac CT examinations could benefit from noise-reducing post-processing with 2D non-linear adaptive filters.

The next step will be to investigate whether noise reduction filters can be used not only to improve the image quality but to actually lower the dose in CCTA, without sacrificing diagnostic accuracy. To answer this question, a more elaborate and careful study design is required, including several levels of dose reduction. A study of this kind is currently being initiated at our center.

Conflict of interest: None.

References

Mettler

, Huda

, Yoshizumo

, Effective doses in radiology and diagnostic nuclear medicine: a catalog. Radiology 2008;248:254–63

Hausleiter

, Meyer

, Hadamitzky

, Radiation dose estimates from cardiac multislice computed tomography in daily practice. Impact of different scanning protocols on effective dose estimates. Circulation 2006;113:1305–10

Weustink

, Molllet

, Neefjes

, Preserved diagnostic performance of dual-source CT coronary angiography with reduced radiation exposure and cancer risk. Radiology 2009;252:53–60

Granlund

GH.

In Search of a General Picture Processing Operator. Comput Graph Image Process 1978;8:155–73

Granlund

, Knutsson

Signal Processing for Computer Vision. Dordrecht: Kluwer Academic Publishers, 1995

Svensson

B. A.

Multidimensional Filtering Framework with Applications to Local Structure Analysis and Image Enhancement (doctoral dissertation). Linköping: Linköping University, 2008

Singh

, Kalra

, Sjöberg

, Application of an adaptive non-linear post processing filter for improving the image quality of low radiation dose CT: A double blinded comparative study. Abstract 1273, American Roentgen Ray Society, ARRS, 2009

Singh

, Kalra

, Forsberg

, Application of noise projection software and 2D non-linear adaptive filters for radiation dose reduction with a renal calculus CT protocol. Abstract 1269, American Roentgen Ray Society, ARRS, 2009

Leander

, Söderberg

, Fält

, Post-processing image filtration enabling dose reduction in standard abdominal CT. Radiat Prot Dosimetry 2010;139:180–5

10.

Ledenius

, Stålhammar

, Wiklund

, Evaluation of image-enhanced paedriatic computed tomography brain examinations. Radiat Prot Dosimetry 2010;139:287–92

11.

Månsson

LG.

Methods for the evaluation of image quality: A review. Radiat Prot Dosimetry 2000;90:89–99

12.

Tingberg

, Herrmann

, Lanhede

, Comparison of two methods for evaluation of the image quality of lumbar spine radiographs. Radiat Prot Dosimetry 2000;90:165–88

13.

Tingberg

, Båth

, Håkansson

, Evaluation of image quality of lumbar spine images: a comparison between FFE and VGA. Radiat Prot Dosimetry 2005;114:53–61

14.

Tingberg

, Herrmann

, Besjakov

, Evaluation of lumbar spine images with added pathology. Proc SPIE 2000;3981:34–42

15.

Börjesson

, Håkansson

, Båth

, A software tool for increased efficiency in observer performance studies in radiology. Radiat Prot Dosimetry 2005;114:45–52

16.

European Commission. CEC European guidelines on quality criteria for computed tomography. EUR 16262 EN. Luxembourg: European Commission, 1999

17.

Smedby

, Fredrikson

Visual grading regression – analysing data from visual grading experiments with regression models. Br J Radiol 2010;83:767–75

18.

Schuetz

, Zacharopoulou

, Schlattmann

, Meta-analysis: Noninvasive coronary angiography using computed tomography versus magnetic resonance imaging. Ann Intern Med 2010;152:167–77

19.

Weustink

, Mollet

, Neefjes

, Diagnostic accuracy and clinical utility of non-invasive testing for coronary disease. Ann Intern Med 2010;152:630–9

20.

Achenbach

, Marwan

, Ropers

, Coronary computed tomography angiography with a consistent dose below 1 mSv using prospectively electrocardiogram-triggered high-pitch spiral acquisition. Eur Heart J 2010;31:340–6

21.

Kak

, Slaney

Principles of computerized tomographic imaging. Piscataway, NJ: IEEE Service Center, 1988

22.

Sunnegårdh

Combining analytical and iterative reconstruction in helical cone-beam CT (doctoral dissertation). Linköping: Linköping University, 2007