Improving image quality with deep learning image reconstruction in double-low-dose head CT angiography compared with standard dose and adaptive statistical iterative reconstruction

Abstract

Objective:

To demonstrate similar image quality with deep learning image reconstruction (DLIR) in reduced contrast medium (CM) and radiation dose (double-low-dose) head CT angiography (CTA), in comparison with standard-dose and adaptive statistical iterative reconstruction-Veo (ASIR-V).

Methods:

A prospective study was performed in 63 patients who under head CTA using 256-slice CT. Patients were randomized into either the standard-dose group (n = 38) with 40 ml of Iopromide (370 mgI ml⁻¹ at 4.5 ml s⁻¹); or a double-low-dose group (n = 25) with CM of 25 ml at 3.0 ml s⁻¹. For image reconstruction, the double-low-dose group used DLIR-M and DLIR-H strength, and the standard-dose group used ASIR-V with 50% strength. The CT value and standard deviation, signal-to-noise ratio and contrast-to-noise ratio of posterior fossa, neck muscles, carotid, vertebral and middle cerebral arteries were measured. The image noise, vessel edge and structure blurring and overall image quality were assessed by using a 5-grade method.

The double-low-dose group reduced CM dose by 37.5% and CT dose index by 41% compared with the standard-dose group. DLIR further reduced the standard deviation value of the middle cerebral artery and posterior fossa and provided better overall subjective image quality (p < 0.05).

Conclusion:

DLIR significantly reduces image noise and provides higher overall image quality in the double-low-dose CTA.

Advances in knowledge

It is feasible to reduce CM dose by 37.5% and volume CT dose index by 41% with the combination of 80 kVp and DLIR in head CTA. Compared with ASIR-V, DLIR further reduces image noise and achieves better image quality with reduced contrast and radiation dose.

Introduction

The number of CT scans has been increasing with the increase of clinical requirements^1,2 and radiation dose reduction has been the focus of many researchers in the last a few decades. However, reducing radiation dose will inevitably increase image noise which will adversely affect the image quality of CT,³ and the observation of subtle lesions.^4,5

There are many ways to reduce image noise and improve image quality under low-dose CT scanning. One of the effective methods is to use iterative reconstruction (IR) algorithms such as the adaptive statistical iterative reconstruction-Veo (ASIR-V) algorithm. In fact, the IR algorithm was used for image reconstruction when the first CT came out, but its development was limited due to its long reconstruction time and the need for expensive computer hardware⁶ in the early days. Thus, filtered back projection (FBP) became the main method of CT image reconstruction due to its fast reconstruction speed for a long time. However, the image noise of FBP images is large in low-dose conditions.

As people demand higher spatial and contrast resolution, lower noise, faster scanning time and lower radiation dose at the same time, new reconstruction techniques have emerged. At present, a variety of reconstruction algorithms using artificial intelligence have been applied to CT images and have demonstrated good image noise reduction capabilities in early clinical applications.^7,8 One of such algorithms is the deep learning image reconstruction (DLIR algorithm developed by GE Healthcare (TrueFidelity^TM, Milwaukee). DLIR features a deep neural network (DNN), which was trained with high quality FBP data sets acquired at high radiation doses to learn how to differentiate noise from signals, and to intelligently suppress the noise without impacting anatomical and pathological structures.

The purpose of our research was to investigate the feasibility of using DLIR to significantly reduce both radiation dose and contrast medium dose while maintaining or improving image quality in head CTA compared with the scans using standard radiation and contrast medium (CM) doses and reconstructed with the state-of-the-art ASIR-V algorithm.

Methods and materials

This was a prospective study. Data were collected from patients who underwent CTA examinations in the First Affiliated Hospital of Xi’an Jiaotong University from July 2020 to September 2020. The procedure followed in this study was in line with the standards set by the hospital medical ethics committee, and the informed consent was signed by the participants. Inclusion criteria: patients who were clinically required to undergo cranial CTA examination. The exclusion criteria were: ① renal insufficiency [Glomerular filtration rate <30 ml/min]; ② inability to perform venipuncture; ③ known allergies; ④ heart failure which will affect the adequate contrast flow due to poor heart function; ⑤ external causes causing image quality too poor to be evaluated (One case with emergency cerebral hemorrhage had severe motion artifacts due to irritability was excluded).

Finally, this study collected 63 patients with an average age of 55.39 ± 13.21 years, range (19–83 years). The patients were divided into a standard dose group and a double low-dose group using a random table method.

Imaging technique and post-processing

All patients underwent head CTA on a 256-slice spiral CT scanner (Revolution CT, GE Healthcare, Milwaukee, USA). All patients were injected with CM (Iopromide, 370 mgI ml⁻¹) using a 20G cannula via median elbow vein, and contrast-enhanced scans were manually triggered. The scan range was from the top of the skull to the second cervical vertebra. The scanning parameters and CM dosage were: for the standard-dose group (n = 38), CM volume, 40 ml, injection rate, 4.5 ml s⁻¹, tube voltage, 100 kVp, noise index (NI), 11.5, resulted in a volume CT dose index (CTDIvol) of 9.67 ± 1.14 mGy Images were reconstructed using ASIR-V with 50% strength (ASIR-V50%); for the double low-dose group (n = 25), CM volume, 25 ml, injection rate, 3.0 ml s⁻¹, tube voltage, 80 kVp, NI, 15, resulted in a CTDIvol of 5.69 ± 0.53 mGy. Images were reconstructed using DLIR with medium (DLIR-M) and high (DLIR-H) strength. Both groups used automatic tube current modulation to meet the NI setting for that group, with an additional 40 ml saline injection after CM at a flow rate of 4.5 ml s⁻¹. Compared with the standard-dose group, the double-low-dose group had a 37.5% reduction in CM volume and a 41% reduction in CTDIvol. The image reconstruction layer thickness was 0.625 mm.

Objective evaluation

The reconstructed images were sent to a GE Advantage Workstation (AW4.7) for measurement and analysis by two general radiologists with 12 and 10 years of experience in CT imaging and diagnosis. Mean CT value and standard deviation (SD) of the posterior fossa, neck muscles, carotid artery, vertebral artery and middle cerebral artery of all subjects were measured to calculate the signal-to-noise ratio (SNR = ROI_vessel/SD_vessel) and contrast-to-noise ratio (CNR) for arteries using muscle as the background: CNR=[(ROI_vessel – ROI_muscle) / SD_muscle], where ROI vessel and ROI muscle represent the average attenuation of the vessel of interest and paravertebral muscle, respectively, and SD muscle represents image noise.⁹ The area of region of interest (ROI) was 50 mm² for the posterior fossa and neck muscle and 2 mm² for the arteries, and ROI was placed in the center of the blood vessel as much as possible during measurement avoiding calcification and plaque areas. Beam hardening, streaks, and/or partial volume artifacts are prone to appear in the posterior fossa due to the large bony structures in this area^10,11 which will artificially increase the non-uniform distribution of CT number or standard deviation. Therefore, in addition to the inherent image noise associated with the scanner and patient-related factors, the SD values in the posterior fossa can also reflect the amount of change in CT numbers caused by beam hardening artifacts.^12–14 However, since DLIR further reduces the overall image noise, to reflect the net change of CT number non-uniformity caused by beam hardening artifacts, we adapted the beam hardening artifact (BHA) index concept introduced by Lin et al¹⁵ to reduce the influence of background image noise of different reconstructions. Specifically, the BHA was defined as: BHA = $\sqrt{{SD}_p^2-{SD}_m^2}$, where SD_p² is the measured SD value for the posterior fossa and SD_m² is the SD value for the neck muscle which was served as background in our study. CT radiation dose, including CTDIvol, dose–length product (DLP), were recorded and the effective radiation dose (effective dose, ED) was calculated using the formula ED = DLP×k, where k = 0.0021 mSv/(mGy·cm) for the head. The noise reduction rate for DLIR was calculated as follows: noise reduction rate (DLIR) (%) = (SD(ASIR-V) – SD(DLIR)) / SD(ASIR-V) × 100, where DLIR indicates DLIR-M or DLIR-H).¹⁴

Subjective image assessment

The same two radiologists used a 5-level scoring method to assess the degree of image noise, blood vessel edge sharpness and structure clarity, and overall image quality. They were blinded to the reconstruction settings and results from the objective image quality analysis by changing the image orders between the objective measurement and subjective rating in two different image reviews. Two general radiologists read the films independently first, any disagreements in scoring were resolved after discussion to reach the final results. The five grades were as follows: image noise (Grade 0, slight; Grade 1, mild; Grade 2, moderate; Grade 3, high; Grade 4, severe) and sharpness and clarity (Grade 0, no blurring; Grade 1, slightly blurred; Level 2, moderately blurred; Level 3, highly blurred, Level 4, severely blurred).¹⁶ The overall image quality of each image was defined as the combination of image noise and blur levels.

Statistical analysis

The SPSS 22 software was used for statistical analysis of the data, and the measurement data were expressed as mean ± standard deviation. Radiation dose-related values between the two groups were compared using the independent sample t-test. The objective measurements (CT value, SD, SNR and CNR) from the two scan groups with three different reconstructions (ASIR50% in the standard dose group and DLIR-M and DLIR-H in the double-low-dose group) were compared using the One-way ANOVA test while the subjective image quality scores were compared using the Kruskal–Wallis test. Cohen’s κ test was used to study the interobserver agreement between the two radiologists for the objective and subjective assessment. A κ value lower than 0.19 was considered poor; 0.20–0.39 was considered fair; 0.40–0.59 was considered moderate; 0.60–0.79 was considered substantial; and 0.80–1.00 was considered perfect. A p < 0.05 in the analysis would indicate that the difference is statistically significant.

Results

In this study, 63 patients who completed head CTA examination were enrolled, including 35 male patients and 28 female patients. They were grouped by random table method and divided into the standard-dose group (n = 38) and double-low-dose group (n = 25). There was no significant difference in the general information of subjects between the two groups (Table 1).

Table 1.

Normal information (objective evaluation)

	Standard-gose group	Double-low-dose group	p
CTDIvol (mGy）	9.67 ± 1.14	5.69 ± 0.53	<0.001
DLP (mGy·cm)	280.78 ± 81.92	183.72 ± 60.96	<0.001
ED (mSv)	0.54 ± 0.27	0.34 ± 0.16	<0.001
CM (ml)	40	25	-
Rate (ml/s)	4.5	3	-
Age (years)	56.62 ± 13.66	53.97 ± 12.74	0.708
Gender (%)
Male	61	48
Female	39	52

CM: contrast medium;CTDIvol: volumetric CT dose index; DLP: dose–length product; ED: effective dose.

Quantitative analysis of the image noise, SNR, CNR and beam hardening artifact index

The mean CT value and SD, SNR and CNR of posterior fossa, neck muscles, carotid artery, vertebral artery, and middle cerebral artery were measured, and BHA index of posterior fossa were calculated. The measurement results of the two radiologists had good agreement with ICC values greater than 0.79 for all measurements. CT values were similar among different reconstruction groups. The background image noise (neck muscle SD values) was reduced with DLIR and further decreased as DLIR level increased. The noise reduction rate further increases with the increase of DLIR level. The noise reduction rates of DLIR-M and DLIR-H in the muscle were 14.03 and 32.59%, respectively. For the SNR and CNR measurements, except for the SNR values of the middle cerebral artery (where DLIR-H and DLIR-M showed better values (p < 0.05)), there was no difference in quantitative measurements among the three groups (p > 0.05). (Table 2, Figure 1). DLIR significantly reduced the BHA in the posterior fossa area with much smaller SD values (Table 2) and BHA index values for DLIR-M and DLIR-H images (p < 0.001) (Figure 2).

Table 2.

Quantitative analysis (objective evaluation)

	Standard-dose group	Double-low-dose group		p
		DLIR-M	DLIR-H
Radiologist 1 carotid artery
HU	387.35 ± 73.18	372.76 ± 95.33	375.61 ± 82.58	0.728
SD	11.89 ± 6.95	12.64 ± 8.86	10.03 ± 8.74	0.759
SNR	52.73 ± 31.47	58.93 ± 25.78	69.32 ± 28.07	0.728
CNR	60.85 ± 32.59	64.48 ± 41.37	81.07 ± 39.49	0.937
Vertebral artery
HU	397.53 ± 83.77	385.75 ± 89.63	385.82 ± 90.03	0.634
SD	11.42 ± 9.25	11.96 ± 8.46	11.07 ± 6.31	0.903
SNR	73.64 ± 75.49	77.35.39 ± 49.27	79.50 ± 49.31	0.877
CNR	64.22 ± 32.65	67.61 ± 43.71	78.94 ± 48.30	0.961
Middle cerebral artery
HU	384.91 ± 67.36	369.79 ± 82.77	369.63 ± 82.07	0.619
SD	15.66 ± 12.08	7.93 ± 3.93	5.44 ± 3.06	＜0.05
SNR	56.26 ± 43.83	106.78 ± 98.70	110.04 ± 134.12	＜0.05
CNR	58.79 ± 38.04	42.94 ± 24.33	48.47 ± 24.90	0.439
Posterior fossa
HU	47.38 ± 5.73	49.84 ± 5.21	49.11 ± 6.07	0.720
SD	21.06 ± 3.56	13.95 ± 2.16	9.49 ± 2.32	<0.001
BHA	18.81 ± 2.73	11.06 ± 2.05	7.02 ± 2.46	<0.001
Neck muscles
HU	52.32 ± 6.09	54.88 ± 7.2 9	56.35 ± 6.27	0.589
SD	9.48 ± 1.36	8.50 ± 2.36	6.39 ± 3.01	0.079
Radiologist 2 carotid artery
HU	381.11 ± 86.54	377.26 ± 83.23	376.61 ± 86.32	0.754
SD	11.54 ± 6.15	12.37 ± 9.74	10.72 ± 8.65	0.728
SNR	55.87 ± 26.65	59.81 ± 26.61	73.28 ± 31.19	0.736
CNR	65.93 ± 33.85	64.94 ± 43.70	77.49 ± 47.73	0.918
Vertebral artery
HU	395.82 ± 82.49	389.26 ± 87.98	389.42 ± 91.15	0.637
SD	12.62 ± 9.12	12.96 ± 8.37	11.59 ± 6.04	0.941
SNR	70.30 ± 79.18	74.04 ± 34.79	79.93 ± 48.48	0.893
CNR	63.38 ± 37.87	67.19 ± 43.89	81.62 ± 46.81	0.972
Middle cerebral artery
HU	376.94 ± 65.38	369.15 ± 78.30	368.25 ± 82.97	0.663
SD	15.09 ± 13.15	6.26 ± 3.49	5.25 ± 3.07	＜0.05
SNR	55.14 ± 47.59	106.46 ± 103.97	113.47 ± 128.74	＜0.05
CNR	52.25 ± 36.59	46.63 ± 22.44	50.31 ± 22.61	0.401
Posterior fossa
HU	45.48 ± 5.27	48.35 ± 5.23	49.17 ± 5.63	0.759
SD	22.16 ± 3.90	12.62 ± 2.63	9.21 ± 1.89	<0.001
BHA	19.98 ± 1.97	9.63 ± 2.37	6.18 ± 2.84	<0.001
Neck muscles
HU	52.20 ± 6.13	54.28 ± 8.37	55.17 ± 6.33	0.594
SD	9.59 ± 1.34	8.16 ± 2.48	6.83 ± 3.26	0.084

Data are mean value ±standard deviation. DLIR-M, and DLIR-H = deep learning reconstruction with medium and high strength levels, respectively; HU = mean CT number, SD = image noise, SNR = ROI_vessel/SD_vessel, and CNR = [(ROI_vessel – ROI_muscle) / SD_muscle],

BHA (Beam hardening artifact) = $\sqrt{{SD}_p^2-{SD}_m^2}$ , where SD_p² is the measured SD value for the posterior fossa and SD_m² is the SD value for the neck muscle.

BHA, beam hardening artifact; CNR, contrast-to-noise ratio; DLIR, deep learning image reconstruction; HU, Hounsfield unit; SD, standard deviation; SNR, signal-to-noise ratio.

Figure 1.

Axial images with 100 kVp ASIR-V50% (a, d, g), 80 kVp with DLIR-H (b, e, h) and DLIR-M (c, f, i) at the second cervical level, median level of orbit, and posterior fossa. DLIR, deep learning image reconstruction.

Figure 2.

Axial images with 100 kVp ASIR-V50% (a), 80 kVp with DLIR-M (b) and DLIR-H (c) at the posterior fossa. ROI was drawn in the interpetrous region (the area indicated by the yellow arrow) to analyze the artifact index. DLIR significantly reduced the beam hardening artifacts in the posterior fossa area. ASIR, adaptive statistical iterative reconstruction; DLIR, deep learning image reconstruction; ROI, region of interest.

Qualitative analysis

The subjective qualitative analysis results for both radiologists are shown in Table 3 and image quality comparison is shown in Figures 1 and 2. The subjective evaluation scores showed that in general as the DLIR intensity level increased, the score gradually decreased from high (worse) to low (better). The observation results of the two radiologists were in substantial agreement (the κ value was 0.658, p < 0.005). Compared with the standard-dose ASIR-V50% images, the overall image quality of the double-low-dose DLIR images were significantly improved between DLIR-H and ASIR-V50% (both radiologists, p < 0.001) (Radiologist 1: 0.15 ± 0.23 vs 2.72 ± 0.58, p < 0.001; Radiologist 2: 0.23 ± 0.35 vs 2.86 ± 0.43, p < 0.001); and between DLIR-M and ASIR-V50% (Radiologist 1: 0.93 ± 0.65 vs 2.72 ± 0.58, p < 0.001; Radiologist 2: 0.83 ± 0.65 vs 2.86 ± 0.43, p < 0.001).

Table 3.

Qualitative analysis

General group Low-dose group
(n = 46)	ASIRV50%	DLIR-M	DLIR-H	p
Radiologist 1
Mottle	2.72 ± 0.58	0.93 ± 0.65	0.08 ± 0.26	<0.001
Blurring	0	0	0.07 ± 0.26	0.019
Overall	2.72 ± 0.58	0.93 ± 0.65	0.15 ± 0.23	<0.001
Radiologist 2
Mottle	2.86 ± 0.43	0.83 ± 0.65	0.12 ± 0.35	<0.001
Blurring	0	0	0.11 ± 0.24	0.021
Overall	2.86 ± 0.43	0.83 ± 0.65	0.23 ± 0.35	<0.001

Data are mean value ±standard deviation.

ASIR, adaptive statistical iterative reconstruction; DLIR, deep learning image reconstruction.

Discussion

The purpose of our study was to compare the image noise and sharpness of the state-of-the-art ASIR and DLIR algorithms in head CTA. We evaluated the DLIR images under reduced radiation dose and contrast medium dose. The results of our study showed that the double-low-dose group using DLIR-M and DLIR-H could further reduce image noise and BHA and achieve better image quality compared with the ASIR-V50% images in the standard-dose group and DLIR images with high intensity level showed the best overall subjective image quality score. The CM dose was reduced by 37.5% and the radiation dose was reduced by 37%.

At present, a large number of studies have shown that thin-slice CT helps to improve the ability to detect lesions.¹⁷ Because the thinner-layer reconstruction in cranial CT angiography will cause an increase in image noise, the image quality will be reduced under the same radiation dose. However, increasing the radiation dose to improve the quality of thin-layer images is against the principle of ALARA (as low as reasonably achievable); at the same time, at the same time, increased use of contrast media also increases the risk of CIN (contrast media nephropathy). Therefore, how to balance radiation dose reduction and image quality is a big challenge. Currently, IR algorithm technology is usually used to reduce radiation dose while ensuring image quality.^18–20

Although the use of IR algorithms can significantly reduce image noise, under very low dose conditions, the image texture and spatial resolution may be altered, especially when using high-intensity IR algorithms, causing “excessive smoothing”, or simply “unnatural looking”.^21,22 Deep-learning techniques have been used in recent years to try to alleviate this problem. Akagi et al⁷ evaluated the use of a deep learning reconstruction in abdominal ultra-high-resolution CT and indicated that the deep-learning based reconstruction algorithm could reduce image noise and improve image quality compared with other IR algorithms under the same scan conditions. Shen et al⁸ proposed a modularized DNN for low-dose CT and demonstrated the competitive performance of this algorithm in reducing noise and maintaining structural fidelity compared with other commercial IR algorithms. DLIR is another deep-learning based reconstruction algorithm that uses artificial intelligence to improve the image quality of CT through deep convolutional neural networks (DNN). These networks optimize the images by minimizing the difference between their output and the ideal training sample images. DLIR is developed to overcome the drawbacks of the conventional IR algorithms. By using the high-dose FBP reconstructions as the gold-standard for training, DLIR can significantly reduce noise amplitude while maintain image sharpness, texture, and structural fidelity. DLIR has been applied clinically²³ and the technical details describing the DLIR algorithm (TrueFidelity^TM) can be found in the manufacturer’s white paper.²⁴ There have been several publications related to the evaluation of DLIR. In our study, we extended the application of DLIR to head where BHA may introduce additional factor to negatively affect image quality and studied the performance of DLIR under the “double low” dose conditions (lower radiation dose and contrast dose). Although previous studies focused on CT images of different organs, our results were still consistent with their results in general. Cao et al²³ pointed out that in contrast-enhanced abdominal CT with extremely low radiation dose, DLIR-H yielded a significantly lower image noise, higher CNR and higher overall image quality than the ASIR-V50% under low signal conditions.²³ Jensen et al²⁵ reported that in abdomen CTs, Truefidelity^TM images showed higher attenuation, lower image noise, and higher CNR values than ASIR-V images with a blending factor of 30%. Park et al²⁶ compared image noise and sharpness of vessels and muscle in lower extremity CT angiography between ASIR-V and DLIR algorithms and found that DLIR-H was the most balanced image in terms of image noise and sharpness among the examined image combinations. Compared with the latest generation IR algorithm (ASIR-V) provided by the same manufacturer, our results showed that the new DLIR algorithm significantly reduced image noise and BHA around the posterior fossa area in head CTA, similar to the results of recently published paper by Kim et al in the routine non-contrast brain CT scans.¹⁴ The low-dose images generated by DLIR had similar spatial resolution and noise texture properties to those of standard-dose ASIR images. On the other hand, DLIR algorithms further reduced the magnitude of image noise, resulted in better overall image quality. Using DLIR to process images greatly benefited the display of low-dose head CT images.

Our research has some limitations. This study had a relatively small number of patients, so the results of the study need further validation. Our research requires further studies with more patients to confirm our findings. In this study, only two radiologists performed subjective image quality assessment. The subjective results of our study need to be evaluated by more neuroradiologists.²⁵ Our research only evaluated the similarity between low-dose scanning using DLIR reconstruction and standard-dose IR; it did not evaluate the disease. The potential diagnostic accuracy improvement for small lesions with low-dose scanning and DLIR needs to be investigated in future research. Our research performed subjective evaluation on spatial resolution but did not explicitly measure it. In the future, modular transfer function and noise power spectrum measurement from phantom and patient studies may be added to have more objective evaluation of how DLIR with different strengths influence it. In addition, whether reducing image layer thickness together with DLIR can improve lesion detection may also be added to the research lists.²⁷

In summary, the application of DLIR technology in cranial CT angiography significantly reduces image noise and further improves overall image quality under lower radiation and contrast doses. DLIR provides clinically acceptable image quality in head CTA with 37% radiation dose reduction and 37.5% contrast medium dose reduction compared with the conventional ASIR-V algorithm using standard radiation and contrast doses.