Application of deep learning image reconstruction algorithm to improve image quality in CT angiography of children with Takayasu arteritis

Abstract

BACKGROUND:

The inflammatory indexes of children with Takayasu arteritis (TAK) usually tend to be normal immediately after treatment, therefore, CT angiography (CTA) has become an important method to evaluate the status of TAK and sometime is even more sensitive than laboratory test results.

OBJECTIVE:

To evaluate image quality improvement in CTA of children diagnosed with TAK using a deep learning image reconstruction (DLIR) in comparison to other image reconstruction algorithms.

METHODS:

hirty-two TAK patients (9.14±4.51 years old) underwent neck, chest and abdominal CTA using 100 kVp were enrolled. Images were reconstructed at 0.625 mm slice thickness using Filtered Back-Projection (FBP), 50%adaptive statistical iterative reconstruction-V (ASIR-V), 100%ASIR-V and DLIR with high setting (DLIR-H). CT number and standard deviation (SD) of the descending aorta and back muscle were measured and contrast-to-noise ratio (CNR) for aorta was calculated. The vessel visualization, overall image noise and diagnostic confidence were evaluated using a 5-point scale (5, excellent; 3, acceptable) by 2 observers.

RESULTS:

There was no significant difference in CT number across images reconstructed using different algorithms. Image noise values (in HU) were 31.36±6.01, 24.96±4.69, 18.46±3.91 and 15.58±3.65, and CNR values for aorta were 11.93±2.12, 15.66±2.37, 22.54±3.34 and 24.02±4.55 using FBP, 50%ASIR-V, 100%ASIR-V and DLIR-H, respectively. The 100%ASIR-V and DLIR-H images had similar noise and CNR (all P > 0.05), and both had lower noise and higher CNR than FBP and 50%ASIR-V images (all P < 0.05). The subjective evaluation suggested that all images were diagnostic for large arteries, however, only 50%ASIR-V and DLIR-H met the diagnostic requirement for small arteries (3.03±0.18 and 3.53±0.51).

CONCLUSION:

DLIR-H improves CTA image quality and diagnostic confidence for TAK patients compared with 50%ASIR-V, and best balances image noise and spatial resolution compared with 100%ASIR-V.

Keywords

Computed Tomography Angiography (CTA)takayasu arteritis pediatrics deep learning image reconstruction

1 Introduction

CT angiography (CTA) is a common examination, which can quickly and noninvasively identify the main artery malformation [1 –6] and has been widely used for children. It is also an important evaluation method for Takayasu arteritis (TAK) [7, 8]. Children with TAK need to use immunosuppressants, the inflammatory indexes of children with TAK usually tend to be normal immediately after treatment, which cannot correctly reflect the severity and activity of the disease [7 , 10]. Therefore, CTA has become an important method to evaluate the status of TAK after surgery and sometime is even more sensitive than laboratory test results. However, image quality of TAK patients need improvement especially under low dose conditions to further increase diagnostic accuracy and to allow potential dose reduction in the future [11].

Iterative reconstruction (IR) algorithms have shown the ability to reduce image noise and have contributed to the radiation dose reduction in CTA for the last decade [12 –15]. However, constrained by the modelling complexity, there is a need to balance spatial resolution and image noise in IR algorithms, and IR with high weights are often limited in clinical applications to avoid blurring artifacts or blotchy appearance in IR images [16, 17], we need new methods to improve the efficiency of reducing radiation dose than IR.

Recently, a deep learning image reconstruction (DLIR) algorithm (GE Healthcare) has been developed based on the artificial intelligence and deep neural networks. The current version of DLIR consists of three different strength settings: low, medium, and high to provide different noise reduction capability. Studies have shown that DLIR at different strengths can further reduce image noise while avoiding the blurring artifacts. As a result, it can improve the image quality as compared with IR algorithms [18 –22]. In our study, we extended the study to pediatric patients with TAK reconstructed with DLIR at the high setting (DLIR-H) to investigate whether DLIR-H could improve image quality under normal scan conditions in comparison with the traditional Filter back projection (FBP) and the state-of-the-art adaptive statistical iterative reconstruction (ASIR-V) algorithm.

2 Materials and methods

2.1 General information

The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). This was a retrospective study approved by the ethics committee of our hospital, and informed consents from patients were waived. The CTA was performed continuously from Nov 1, 2019 to Jun 30, 2020. Inclusion criteria: 1. clinical diagnosis of TAK; 2. CT examination was needed to determine whether there is arterial stenosis or follow up CT after treatment. Exclusion criteria: 1. there was a history of iodine containing contrast medium allergy; 2. Patients with age > 18 years old.

2.2 CT scan and image reconstruction

All examinations were performed on a 256 rows multi-detector CT scanner (Revolution CT, GE Healthcare, USA), using a tube voltage of 100 kVp, helical pitch value of 1.375 : 1 with 40 mm detector width and 0.5s rotation speed. The automatic tube current modulation technique was used with tube current in the range of 50–500 mA, noise index was 11. All scans were performed when children were in a quiet state. For those children who could not cooperate, the 10%chloral hydrate at a dose of 0.4 ml/kg based on body weight was given orally and scans would not start until patients fell asleep. The scanning range included neck, chest, and abdomen, with the upper edge from the skull base and the lower edge to the anterior superior iliac spine.

For the contrast-enhanced CT protocol, a peripheral venous cannula was pre-placed in the superficial vein of dorsum of the hand. An iodinated contrast agent (320 mg I/ml iodixanol; GE Healthcare, USA) was administered using a single-head power injector. The contrast medium dosage was calculated according to the body weight of each child: 1.6 ml/kg for children lower than 16.0 kg, 1.4 ml/kg for 16.1–35.0 kg and 1.2 ml/kg for higher than 35.0 kg (the maximum usage limit was 70 ml). Flow rate was adjusted according to a fixed injection time of 15s and contrast enhanced scan started at 17s after the start of contrast injection.

The raw data was reconstructed to 0.625 mm thin slice images using 3 types of reconstruction algorithms into 4 groups: FBP group, 50%ASIR-V group, 100%ASIR-V group, and DLIR-H group. All reconstructions used a standard kernel.

2.3 Image quality evaluation

All images were transferred to an advantage workstation (AW4.7, GE Healthcare, USA). Two observers with 15 and 8 years of experience in reviewing CTA images subjectively evaluated image quality using a 5-point scoring system. The reviewers were aware of the clinical task of diagnosing TAK but were blinded to patient information and reconstruction algorithms. The readers evaluated all 4 differently reconstructed images of the same patient before moving to different patients, but the order in which the different reconstructions appear on the AW changed randomly. The observers could freely adjust the window width and position for observation and used multiplanar reconstructions (MPR) and 3-dimensional images for the evaluation. If the scores given by the two observers were not the same, a third senior doctor with 20 years of experience in reviewing children chest CT images would evaluate the images and gave the final score. Subjective scores included the overall image noise, visualization of large arteries (artery diameter greater than the vertebral artery or proper hepatic artery), visualization of small arteries (artery diameter less than or equal to the vertebral artery or proper hepatic artery) and diagnostic confidence. Scores with at least 3 points were accepted for diagnosis with 5 points being the best. The specific evaluation criteria are listed in Table 1.

Table 1
Specific criteria for subjective evaluation

1 point 2 points 3 points 4 points 5 points

Overall image noise severe high moderate little rare

Vessel display ability non-detect detectable measurable clear clear with sharp edges

Diagnostic confidence none little moderate high definite

	1 point	2 points	3 points	4 points	5 points
Overall image noise	severe	high	moderate	little	rare
Vessel display ability	non-detect	detectable	measurable	clear	clear with sharp edges
Diagnostic confidence	none	little	moderate	high	definite

After the subjective evaluation, the two observers conducted objective measurement on the AW workstation together, selected the largest section of the liver portal, set a circular region of interest (ROI) on the descending aorta (Ao) on the liver portal section with a diameter half of that of Ao and on the back muscle (Mu) at the same imaging level to measure their CT attenuation value and standard deviation (SD) value. The contrast-to-noise ratio (CNR) for the descending aorta was calculated using the following formula: $CNR = (CT (Ao) - CT (Mu)) / ((SD (Ao) + SD (Mu)) / 2)$

2.4 Statistical analysis

All the data were expressed using the form of mean±standard deviation. The differences of the 4 image groups were analyzed, continuous variables following the normal distribution were analyzed by using the repeated measures analysis of variance with Bonferroni correction. The ordinal scales or variables that failed to follow normal distribution were analyzed by using Friedman’s test. P < 0.05 was considered to have statistically significant difference. The agreement for the subjective analysis between the two readers was evaluated by using the Kappa test, Kappa values (k) < 0 were considered as no agreement, 0.0 < k = < 0.2 as poor, 0.2 < k = < 0.4 as fair, 0.4 < k = < 0.6 as moderate, 0.6 < k = < 0.8 as substantial, and 0.8 < k = < 1.0 as excellent.

3 Results

A total of 32 patients (9 males and 23 females) with an average age of 9.14±4.51 years (1–17 years) were included in the study. The volume CT dose index (CTDIvol) for the group was 4.53±2.62 mGy and the dose-length-product (DLP) was 258.39±189.93 mGy.cm.

The results of the subjective evaluation and objective measurement are listed in Tables 2 3, respectively. For the subjective score, the overall image noise of the 50%ASIR-V, 100%ASIR-V and DLIR-H images could meet the diagnostic requirements, with the 100%ASIR-V and DLIR-H images having the highest scores, and there was no statistical difference between them (p = 0.10); for the visualization of large arteries, all images could met the diagnostic requirements (Fig. 1), and 100%ASIR-V and DLIR-H images were the best, and there was no statistical difference between them (p = 0.57); for the visualization of small arteries, the 100%ASIR-V over-smoothed images with sub-optimal resolution (2.84±0.37) and only the 50%ASIR-V and DLIR-H could provide the satisfactory diagnostic confidence for all vessels, with DLIR-H being the best reconstruction algorithm (Fig. 2); The DLIR-H provided significantly better diagnostic confidence (4.09±0.30) than that of 50%ASIR-V (3.03±0.18) and 100%ASIR-V (3.00±0.00) (p < 0.001). The agreement between the two readers was substantial with respect to the overall image quality (k = 0.81, p < 0.05).

Table 2
Results of subjective scores

FBP 50%ASIR-V 100%ASIR-V DLIR P value

Overall image noise 2.16±0.37 3.09±0.30 4.41±0.50^* 4.03±0.18^* < 0.001

Display ability of large arteries 3.06±0.25 4.03±0.31^* 4.59±0.50^# 4.34±0.48^^# < 0.001

Display ability of small arteries 2.94±0.25^^# 3.03±0.18^^† 2.84±0.37^#^† 3.53±0.51 < 0.001

Diagnostic confidence 2.91±0.30^^# 3.03±0.18^*^† 3.00±0.00^#^† 4.09±0.30 < 0.001

	FBP	50%ASIR-V	100%ASIR-V	DLIR	P value
Overall image noise	2.16±0.37	3.09±0.30	4.41±0.50^*	4.03±0.18^*	< 0.001
Display ability of large arteries	3.06±0.25	4.03±0.31^*	4.59±0.50^#	4.34±0.48^*^#	< 0.001
Display ability of small arteries	2.94±0.25^*^#	3.03±0.18^*^†	2.84±0.37^#^†	3.53±0.51	< 0.001
Diagnostic confidence	2.91±0.30^*^#	3.03±0.18^*^†	3.00±0.00^#^†	4.09±0.30	< 0.001

^*^# †: without in-group statistical difference.

Table 3

Results of objective scores

	FBP	50%ASIR-V	100%ASIR-V	DLIR	P value
Artery CT value^*	391.27±55.23	391.92±54.48	392.28±53.79	393.61±54.19	1.00
Artery SD value	31.36±6.01	24.96±4.69	18.46±3.91^*	15.58±3.65^*	< 0.001
Muscle CT value^*	60.46±9.38	60.42±8.74	60.22±8.10	60.54±7.69	1.00
Muscle SD value	24.79±5.16	17.68±3.52	11.22±2.40^*	12.64±2.71^*	< 0.001
Artery CNR	11.93±2.12	15.66±2.37	22.54±3.34^*	24.02±4.55^*	< 0.001

CNR: contrast to noise ratio. ^*: without in-group statistical difference.

Fig. 1

The 3D reconstruction images of a 14-years old girl suffered from Takayasu arteritis. The scan voltage was 100 kV, 1A-1D was the 0.625 mm image with FBP, 50%ASIR-V, 100%ASIR-V and DLIR-H. respectively. The image noise of 1C and 1D were lower than that of 1A and 1B (×area). All images could show the slight dilation of big arteries (white arrow), and the stenosis of right renal artery (black arrow), 1C and 1D images had the least image noise.

Fig. 2

The multi-planar-reformat (MPR) images of the same 14-years old girl in Fig. 1. 2A-2D was the 0.625 mm image with FBP, 50%ASIR-V, 100%ASIR-V and DLIR-H, respectively. The 100%ASIR-V image (2C) had the least image noise but was also blotchy in the image texture (black arrow), the boundary of muscles was unclear compared to 2D. The margin of the small artery was blurred and not consistent (white arrow), which could induce the misdiagnosis as vascular occlusion. DLIR-H image (2D) had similar low image noise as 2C but without the pixilated appearance seen in 2C and provided the highest diagnostic confidence for the vessels.

For the objective measurement results: there was no statistical difference in the vascular CT number and muscle CT number among the reconstruction groups; for the vascular noise and muscle noise, the values with 100%ASIR-V (18.46±3.91HU and 11.22±2.40HU) and DLIR-H (15.58±3.65HU and 12.64±2.71HU) had no statistically significant difference between them (p = 0.72 and 0.88), and both significantly lower those (24.96±4.69HU and 17.68±3.52HU) of 50%ASIR-V images. For CNR, the values with 100%ASIR-V (22.54±3.34) and DLIR-H (24.02±4.55) also had no statistically significant difference between them (p = 0.73) and were both significantly higher than that (15.66±2.37) with 50%ASIR-V (p < 0.05). Compared with the FBP images, DLIR-H images reduced image noise by 50.32%in vessels and 49%in soft tissue, and the 50%ASIR-V reduced images noise by 20.41%in vessels and 28.68%in soft tissue; For CNR, DLIR-H increased the value by 101.34%and 50%ASIR-V by 88.94%.

4 Discussion

In CT reconstruction, FBP is the original reconstruction algorithm, the 50%ASIR-V is the recommended iterative reconstruction algorithm currently used in our daily clinical practice, and the images reconstructed with 50%ASIR-V were read and interpreted by radiologists in the original clinical practice to diagnose diseases. On the other hand, 100%ASIR-V has the strongest noise reduction capability for the ASIR-V algorithm, and study has shown that DLIR-H provided the excellence balance between maintaining image spatial resolution and reducing image noise in DLIR images [21]. Therefore, these 4 algorithms were selected for comparing image quality of TAK patients.

In our study we investigated whether DLIR-H could improve image quality in CTA of pediatric patients with TAK under normal scan conditions, in comparison with the traditional FBP and the state-of-the-art 50%ASIR-V algorithm. From the objective evaluation point of view, our results showed that DLIR-H provided about 50%image noise reduction, while 50%ASIR-V provided about 25%reduction compared with FBP. From the subjective evaluation point of view, our results showed that DLIR-H did not affect the visualization of small arteries and image texture while reducing image noise. The overall appearance of DLIR-H images was similar to that of FBP and 50%ASIR-V images, without “blotchy” feeling, which was conducive to the display of detailed structures such as small arteries. In fact, since the overall image noise was significantly reduced, the ability of displaying small vessels was judged to be even better than that of FBP and 50%ASIR-V (3.53 vs. 2.94 and 3.03). All the subjective image quality scores of DLIR-H were higher than 3.0, and its comprehensive score was better than other reconstruction algorithms. On the other hand, even though 100%ASIR-V also significantly reduced image noise, it also altered image texture making images appeared too “smooth” with some spatial resolution and structure detail loses (Fig. 2), which affected the subjective diagnostic confidence and the diameter measurement for small arteries.

CTA is an important evaluation method for TAK [18, 19]. Since TAK can involve the whole artery system, the scanning range is general much larger (including neck, chest and abdomen simultaneously) than that of conventional CTA. In addition, it is also necessary to observe whether the arterial branches with small arteries are involved, high spatial resolution and thin-layer images are desirable. All these clinical requirements put more pressure on the radiation dose in CTA for TAK patients [23], and efficient solutions are urgently needed to reduce the radiation dose while maintaining or improving image quality.

IR algorithms have been widely used to tackle the problem of higher image noise with thin-layer images and/or under reduced radiation conditions. However, IR algorithms with high weights tend to over smooth the high frequency noise and reduce the average frequency of the noise power spectrum (NPS) of the images. Although the image noise is greatly reduced, images may lose the normal image texture (Fig. 2), making images look “overly smooth”, “plastic”, or simply “unnatural” [16 , 24]. Our results indicated that although it did not affect the visualization of large arteries, it was not conducive to display image details such as small arteries using 100%ASIR-V algorithm. Therefore, in general, 50%weight of IR is recommended to maintain the balance between noise and image resolution in clinical routine work. Only when the radiation dose is very low, 100%IR weighted is recommended as a mean to salvage the scans [25]. This is the major limitation of low-dose CTA applicated in TAK.

DLIR is a new generation of image reconstruction algorithm. The deep learning method is used to analyze the characteristics of image noise generation and distribution in different organizations. In theory, the targeted reduction of noise will have minimum impact on the image resolution. In the process of DLIR reconstruction, the integrity of the image information remains largely unchanged when the image noise is greatly reduced, and there is no artificial addition or loss of image resolution.

Our results indicated that when DLIR algorithm was used in the assessment of TAK, the image details were maintained, while image noise was significantly reduced. Ever though our study was focused on image noise reduction and image quality improvement, since at the current dose level with 50%ASIR-V, image quality was acceptable, the lower image noise and better image quality performance of DLIR-H could be traded with radiation dose reduction in the future.

There were still some limitations in this study. Firstly, the case number in the study was rather small, because TAK is not a common disease, so it was hard to collect large samples, and impossible to compare children of different ages in groups. In the future, the sample size will be increased to enrich the data information. Secondly, subjective image evaluation was performed using a non-validated scale developed by the authors. Third, our study focused on the image quality improvement over the state-of-the-art ASIR-V algorithm using the same scan under the standard-dose condition. Although the image quality of some ASIR-V images was different from those of DLIR images, no children had interventional radiology results as the gold standard, so it was impossible to confirm the diagnostic accuracies of the different reconstruction algorithms. Further research is needed to demonstrate the improvement in diagnostic efficacy. In addition, further research is also needed to focus on dose reduction to investigate whether DLIR could achieve the same image quality and diagnostic efficacy as the standard-dose ASIR-V images under lower-radiation dose or lower-contrast agent dose.

5 Conclusion

DLIR-H improves the CTA image quality and diagnostic confidence for TAK patients compared with the state-of-the-art 50%ASIR-V, and best balances image noise and spatial resolution compared with 100%ASIR-V.

6 Funding

This study was funded by Beijing Children’s Hospital Young Investigator Program (grant number BCH-YIPB-2016-06).

References

Tola

, Ozturk

, Yildiz

, et al., Assessment of children with vascular ring, Pediatr Int 59 (2017), 134–140.

Sun

J.H.

, Li

H.Y.

, Yang

L.X.

, et al., Application of 70kVp in abdominal CT angiography to reduce both radiation and contrast dosage and improve patient comfort for children, J Xray Sci Technol 29(5) (2021), 813–821.

Trotman-Dickenson

Congenital lung disease in the adult: guide to the evaluation and management, J Thorac Imaging 30 (2015), 46–59.

Kim

H.J.

, Shin

K.E.

, Park

J.S.

, et al., Intralobar pulmonary sequestration with cystic degeneration mimicking a bronchogenic cyst in an elderly patient: A case report and literature review, Medicine (Baltimore) 99 (2020), e19347.

Zhang

, Zeng

, Chen

, et al., Distribution, diagnosis, and treatment of pulmonary sequestration: Report of 208 cases, J Pediatr Surg 54 (2019), 1286–1292.

Hermelijn

C.M.

, Elders

B.B.

, Ciet

, et al., A clinical guideline for structured assessment of CT-imaging in congenital lung bnormalities, Paediatr Respir Rev 20 (2020), S1526-0542(20)30019-1.

Duftner

, Dejaco

, Sepriano

, et al., Imaging in diagnosis, outcome prediction and monitoring of large vessel vasculitis: a systematic literature review and meta-analysis informing the EULAR recommendations, RMD Open 4 (2018), e000612.

Bardi

, Diamantopoulos

A.P.

EULAR recommendations for the use of imaging in large vessel vasculitis in clinical practice summary, Radiol Med 124 (2019), 965–972.

Direskeneli

, Aydin

S.Z.

, Merkel

P.A.

Assessment of disease activity and progression in Takayasu’s arteritis, Clin Exp Rheumatol 29 (2011), S86–91.

10.

Mukhtyar

, Guillevin

, Cid

M.C.

, et al., EULAR recommendations for the management of large vessel vasculitis, AnnRheum Dis 8 (2019), 318–323.

11.

The ALARA (as low as reasonably achievable) concept in pediatric CT intelligent dose reduction. Multidisciplinary conference organized by the Society of Pediatric Radiology. August 18-19, 2001, Pediatr Radiol 32 (2002), 217–313.

12.

Schindera

S.T.

, Diedrichsen

, Müller

H.C.

, et al., Iterative reconstruction algorithm for abdominal multidetector CT at different tube voltages: Assessment of diagnostic accuracy, Image quality, and radiation dose in a phantom study, Radiology 260 (2011), 454–462.

13.

Goenka

A.H.

, Herts

B.R.

, Dong

, et al., Image noise, CNR, and detectability of low-contrast, low-attenuation liver lesions in a phantom: Effects of radiation exposure, phantom size, integrated circuit detector, and iterative reconstruction, Radiology 280 (2016), 475–482.

14.

Yan

, Xu

, Liang

, et al., Radiation dose reduction by using CT with iterative model reconstruction in patients with pulmonary invasive Fungal infection, Radiology 288 (2018), 285–292.

15.

Harun

H.H.

, Karim

M.K.A.

, Abbas

, et al., Effect of iterative reconstruction algorithm levels on noise index and figure-of-merit in CT pulmonary angiography examinations, J Xray Sci Technol 28 (2020), 893–903.

16.

Greffier

, Frandon

, Larbi

, et al., CT iterative reconstruction algorithms: a task-based image quality assessment, Eur Radiol 30 (2020), 487–500.

17.

Mileto

, Guimaraes

L.S.

, McCollough

C.H.

, et al., State of the art in abdominal CT: The limits of iterative reconstruction algorithms, Radiology 293 (2019), 491–503.

18.

Benz

D.C.

, Benetos

, Rampidis

, et al., Validation of deep-learning image reconstruction for coronary computed tomography angiography: Impact on noise, image quality and diagnostic accuracy, J Cardiovasc Comput Tomogr 13 (2020), S1934–5925(19)30464-2.

19.

Akagi

, Nakamura

, Higaki

, et al., Deep learning reconstruction improves image quality of abdominal ultra-high-resolution CT, Eur Radiol 29 (2019), 6163–6171.

20.

Tatsugami

, Higaki

, Nakamura

, et al., Deep learning-based image restoration algorithm for coronary CT angiography, Eur Radiol 29 (2019), 5322–5329.

21.

Greffier

, Hamard

, Pereira

, et al., Image quality and dose reduction opportunity of deep learning image reconstruction algorithm for CT: a phantom study, Eur Radiol 30 (2020), 3951–3959.

22.

L.L.

, Wang

, Song

, et al., A feasibility study of realizing low-dose abdominal CT using deep learning image reconstruction algorithm, J Xray Sci Technol 29 (2021), 361–372.

23.

Masuda

, Funama

, Nakaura

, et al., Radiation dose reduction with a low-tube voltage technique for pediatric chest computed tomographic angiography based on the contrast-to-noise ratio index, Can Assoc Radiol J 69 (2018), 390–396.

24.

Geyer