Iterative reconstruction vs deep learning image reconstruction: comparison of image quality and diagnostic accuracy of arterial stenosis in low-dose lower extremity CT angiography

Abstract

Objective

To compare image quality and diagnostic accuracy of arterial stenosis in low-dose lower-extremity CT angiography (CTA) between adaptive statistical iterative reconstruction-V (ASIR-V) and deep learning image reconstruction (DLIR) algorithms.

Methods

46 patients undergoing low-dose lower-extremity CTA were enrolled. Images were reconstructed using ASIR-V (blending factor of 50% (AV-50) and 100% (AV-100)) and DLIR (medium (DL-M), and high (DL-H)). CT values and standard deviation of the aorta, psoas, popliteal artery, popliteal and ankle muscles were measured. The edge-rise distance and edge-rise slope were calculated. The degrees of granularity and edge blurring were assessed using a 5-point scale. The stenosis degrees were measured on the four reconstructions, and their mean square errors against that of digital subtraction angiography were calculated and compared.

Results

For both ASIR-V and DLIR, higher reconstruction intensity generated lower noise and higher signal-to-noise ratio and contrast-to-noise ratio values. The standard deviation values in AV-100 images were significantly lower than other reconstructions. The two DLIR image groups had higher edge-rise slope and lower edge-rise distance (DL-M:1.79 ± 0.37 mm and DL-H:1.82 ± 0.38 mm vs AV-50:1.96 ± 0.39 mm and AV-100:2.01 ± 0.36 mm, p = 0.014) than ASIR-V images. The overall image quality of DLIR was rated higher than ASIR-V (DL-M:0.83 ± 0.61, DL-H:0.41 ± 0.62, AV-50:1.85 ± 0.60 and AV-100:2.37 ± 0.77, p < 0.001), with DL-H having the highest overall image quality score. For stenosis measurement, DL-H had the lowest mean-square-errors compared to digital subtraction angiography among all reconstruction groups.

Conclusion

DLIR images had higher image quality ratings with lower image noise and sharper vessel walls in low-dose lower-extremity CTA, and DL-H provides the best overall image quality and highest accuracy in diagnosing artery stenoses.

Advances in knowledge:

DLIR provides high-quality images with sharper edges compared to ASIR-V during low-dose CTA of lower extremity arteries, and DLIR (high) provides the best overall image quality and highest accuracy in diagnosing artery stenoses among all reconstruction algorithms (ASIR-V and DLIR). ASIR-V with blending factor of 100% has the strongest noise reduction ability among all reconstruction algorithms (ASIR-V and DLIR); however, it generates the most blurred images.

Introduction

CT angiography (CTA) of lower extremity arteries is currently the most widely used clinical examination modality for lower extremity arterial diseases. It has a sensitivity of 95% and a specificity of 96% for the diagnosis of more than 50% of the lower extremity arterial lumen stenosis, whose diagnostic performance is basically equivalent to the gold-standard digital subtraction angiography (DSA) examination. ¹ However, CTA of lower extremity arteries also has certain problems, among which the high radiation dose caused by large scanning range (usually larger than 100 cm) is a significant problem. Lower extremity arterial scanning with reduced radiation dose has been a frequent area of interest of imaging scholars. However, as we all know, the reduction of radiation dose will inevitably lead to the loss of image quality, and excessive noise is likely to affect the accurate diagnosis of vascular plaque and other diseases. Therefore, it is vitally important to reduce the radiation dose while ensuring the image quality, so as not to affect the accurate diagnosis of the disease.

Adaptive statistical iterative reconstruction-V (ASIR-V) is a hybrid model-based iterative reconstruction platform developed by GE healthcare, which can reduce image noise and maintain image quality during low-dose scanning. ² The blending factor of ASIR-V can be selected from 10 to 100% according to different clinical indication, image quality and radiation dose requirements. The higher the percentage of the selected iteration level, the lower the image noise. However, some studies have pointed out that the use of high iteration level may produce waxy artifacts in the process of reducing image noise, which will blur the images and affect the accurate diagnosis of lesions to some extent. ^3,4 This adverse effect will become more prominent as the selected blending factor increases.

The deep learning-based reconstruction method is a new type of image reconstruction algorithm, which can reduce image noise without compromising image resolution when compared with the conventional iterative reconstruction algorithm. ⁵ One of such reconstruction algorithms is the deep learning image reconstruction (DLIR) algorithm developed by GE Healthcare which is based on a deep convolutional neural network to simulate the texture of standard-dose filtered back projection (FBP) images which in theory can provide powerful noise reduction capabilities while maintaining high spatial resolution of detailed structures. ⁶ The DLIR algorithm uses a deep neural network, (DNN) that can handle millions of parameters and the training process uses thousands of phantom and patient image sets that were acquired at both the standard and low radiation dose levels. The data sets that were acquired at the standard radiation dose were reconstructed with the FBP algorithm and corrected for many system imperfections such as beam hardening and scattering and are basically free of artifacts while the low dose data sets were reconstructed using DLIR algorithm. During training, these two images are compared across multiple parameters such as image noise, low contrast resolution, low contrast detectability, noise texture, etc. The DNN analyzes the data, minimizes the differences between the two image sets, and synthesizes the reconstruction function (the inference engine), which is optimized through the learning process. The inference engine is then used to validate the extended test data set. This vendor-specific version of DLIR currently works in single-energy-mode CT, allowing the selection of low, medium and high reconstruction strengths, which translates into the degree of desired noise reduction. ^7,8 The purpose of this paper was to validate the hypothesis that DLIR can significantly reduce image noise and improve the image quality and to further evaluate the diagnostic performance of detecting lumen stenosis compared with the adaptive statistical iterative reconstruction-V (ASIR-V) algorithm in the low-dose CTA scan of the lower extremity arteries.

Methods and materials

General information of the patients

This study was a prospective study approved by our institutional ethics committee, and all subjects signed informed consent before the study. The patients who underwent lower extremity arterial CTA from July 2020 to January 2021 in our hospital (the First Affiliated Hospital of Xi’an Jiaotong University) were continuously collected. The inclusion criteria were: clinical application for lower extremity arterial CTA examination, and agreed to be the subject of this study. The exclusion criteria are: (1) known allergy to iodine; (2) renal insufficiency (glomerular filtration rate <30 ml/min/s); (3) failure to complete the examination due to inability to immobility or other reasons. Age, gender, height, weight and other general information of all patients were recorded.

Scanning and injection parameters

All subjects underwent CTA scanning of lower extremity arteries using a 256-row Revolution CT machine (GE Healthcare, Milwaukee). Subjects underwent antecubital vein puncture with a 20-gauge trocar before the scan and were informed to keep strictly immobile during the scan. An automatic bolus-tracking threshold-triggering method was used to determine the delay time after the contrast medium injection. The monitoring level for the bolus tracking scans was located at the upper edge of the fourth lumbar vertebral body. CTA scan started 12 s after the preset threshold attenuation of 150 HU was reached.

The scanning coverage ranged from the upper edge of abdominal aortic bifurcation (the upper edge of the fourth lumbar vertebra) to the toe. Other scanning parameters were: tube voltage of 80 kVp, tube currents of 10–600 mA with Smart mA technique, noise index (NI) of 16 HU. Iohexol (350 mgI ml⁻¹) was used as the contrast agent, and a double-barreled high-pressure syringe was used in a segmented injection mode: First, 50 ml contrast agent was injected at a flow rate of 4 ml s⁻¹, and then 40 ml contrast agent was injected at a flow rate of 2.5 ml s⁻¹ followed by an immediate 40 ml saline flush at a rate of 2.5 ml s⁻¹.

Digital subtraction angiography procedure

The conventional digital subtraction angiography (DSA) was also performed within 2 weeks after CTA. The DSA examination and diagnosis of all patients in our study were carried out by two peripheral vascular doctors with more than 10 years of experience in our hospital with consensus, and the examination was carried out in the interventional operation room by using local angiography on a Philips FD20 angiograph (FD20, Philips Corporation). At least two different views were obtained for each main vessel. The contrast agent with 370 mgI ml⁻¹ concentration (Iopamidol, Bracco) was injected with a high-pressure syringe (PRM, Medrad inc.) at a flow rate of 3.0–5.0 ml s⁻¹. 25–35 ml undiluted contrast medium was used for each position. The lesions were evaluated, their stenotic degrees reported, and images were collected during the operation. If patient was in need for endovascular therapeutic intervention, procedure would be taken in the same period of time.

Image reconstruction and analysis

The original low-dose scan data were reconstructed into four different axial image sets using different reconstruction algorithms, namely: ASIR-V with iterative intensity of 50% (AV-50) and 100% (AV-100) and DLIR with medium (DL-M) and high (DL-H) intensity level. The images reconstructed using the standard ASIR-V at 50% iterative intensity (AV-50 images) were used as reference standards for comparison. The image analysis included both objective and subjective evaluation.

Quantitative analysis

Region of interest (ROI) was positioned on the arterial lumen, muscle and subcutaneous fat at the abdominal aorta level and popliteal artery level and the muscle and fat at the ankle joint layer of the lower limb arterial images of the four reconstruction methods to measure the CT value and standard deviation (SD value). ROI was placed at the center of the arterial lumen avoiding calcification or plaque areas for lumen measurements and perivessel tissue with uniform density during muscle and subcutaneous fat measurement to avoid interference with the measurement results. The area of all ROIs was adjusted according to the area of the measurement site. Calculate the signal-to-noise ratio (SNR, calculation method: Target CT value/Target SD value) and contrast-to-noise ratio (CNR, calculation method: (Target CT value-fat CT value)/ Target SD value) of the vascular lumen and muscles at the measurement level.

The image spatial resolution was evaluated using the sharpness of vessel edges, which was calculated using the ImageJ software (National Institutes of Health, Bethesda, MD) (http://rsb.info.nih.gov/ij). First, we selected the straight section of the lower extremity blood vessel and draw a straight line through the lumen from the fat around the lumen on the transverse section of the blood vessel and use the “plot profile” tool in the “analysis” tab to generate the profile curve between the spatial location and its CT value on the straight line. Again, calcifications and plaques were carefully avoided. This CT attenuation profile was generated at precisely the same location for images reconstructed with ASIR-V and DLIR algorithms. The X axis of the curve represents the spatial location, and the Y axis represents the CT value, and a data table of the X- and Y values of each point on this curve was obtained. This operation was repeated four more times by shifting the straight line up and down along the vessel to get five data sets, and the five data sets were averaged to get a final vessel profile. The edge-rise distance (ERD) and edge-rise slope (ERS) of the vessel profile were then calculated to reflect the sharpness of vessel lumen in the images. ⁹ ERD was calculated as: X1 (X value corresponding to 90% of the maximum Y value (Y1)) minus the X2 (X value corresponding to 10% of the maximum Y value (Y2)) of the rising section of the curve. ERS was calculated as: (Y1-Y2)/ERD. ¹⁰ Smaller ERD (larger ERS) represents sharper edges; and larger ERD (smaller ERS) represents more blurred edges. ERD and ERS were measured and calculated for all reconstructed images of all patients. The diagram for calculating ERD and ERS is shown in Figure 1.

Figure 1.

Schematic diagram of measurement and calculation of ERS and ERD (a) shows the measurement method for measuring ERD and ERS. The two red circles on the figure refer to the ROI for measuring SD values on muscle and fat. (b) lines 1, 2, 3, 4 and 5 in the figure are the curves of five consecutive lines drawn by us, and A is the average curve of these five lines. (c) shows the schematic diagram of ERD and ERS calculation. (d) A 67-year-old male patient, CT attenuation-distance curves obtained at the same level for the four different reconstruction groups. ERD, edge-rise distance; ERS, edge-rise slope; ROI, region of interest; SD, standard deviation.

Qualitative evaluation

Two radiologists with 8 and 9 years of peripheral blood vessel image diagnosis experience independently scored the overall image quality, including image granularity and edge blurring, of all images subjectively using a 5-point scale. ⁷ The subjective image quality evaluation was defined in the following way. For image granularity: score 0, diffuse homogeneous without granularity; score 1, mild image granularity; score 2, moderate image granularity; score 3, substantial image granularity; and score 4, diffuse inhomogeneous with severe granularity. For edge blurring: score 0, no or minimal blurring with crisp edges and well-defined margins; score 1, minor blurring with good edge definition and easily discernable margins; score 2, moderate blurring with slightly poor edge definition and discernable margins; score 3, substantial blurring with poor edge definition and difficultly discernable margins, and score 4, severe blurring with very poor edge definition and indiscernible margins. In order to ensure the consistency of the evaluation results, two radiologists received unified training before image evaluation on how to evaluate image granularity and blurring using lower-extremity CTA images that were not included in this study.

Evaluation of diagnostic accuracy

The vascular analysis software of GE Advantage workstation v. 4.7 (AW4.7, GE Healthcare, Wisconsin) was used for measuring the stenosis rate of all patients with greater than 50% arterial lumen stenosis of the four different reconstructed images. After manually importing the image and tracing out the stenotic vessel segments, which was selected for automatic machine measurement with the area percentage mode selected. Finally, vascular stenosis rates measured on four groups of CTA images were compared with that of DSA, and the mean square error (MSE) between the CT and DSA results was used to characterize the accuracy of the measurement of vascular stenosis rate.

Statistical analysis

Statistical analysis was performed using SPSS software Windows v. 20 (IBM, Chicago, Illinois). The measurement data (CT value, SD value, SNR, CNR, ERD, ERS, granularity, blurring) were all expressed as mean ± SD. The continuous measurements of multiple groups were tested by repeated measure ANOVA with the Bonferroni post-hoc test, and the pairwise comparison of multiple groups of measurement data was tested by LSD. Image quality scores (granularity, blurring and overall) among the four reconstruction groups were tested using Friedman’s test separately for the two observers. p < 0.05 indicates that the difference is statistically significant. Bland–Altman plot was used to evaluate the consistency of the four reconstruction methods with DSA in the measurement of lower extremity arterial stenosis. For the evaluation of the agreement between the two observers in the qualitative image quality evaluation, the weighted κ statistic was used. A κ statistic in the range of 0.81–1.00 was interpreted as excellent, 0.61–0.80 as substantial, 0.41–0.60 as moderate, 0.21–0.40 as fair, and 0.00–0.20 as poor agreement.

Results

According to the inclusion and exclusion criteria, 46 patients aged 29–87 years old, with an average age of (62.37 ± 15.12) years were finally enrolled, of which 37 were males (80.43%, aged 29–87 years, with an average of (61.70 ± 15.36) years old). All patients underwent low-dose lower extremity CTA examination. The average CT dose index (CTDI_vol), dose–length product (DLP) and the effective dose (ED) was (1.09 ± 0.37) mGy, (148.44 ± 49.36) mGy-cm, and (2.08 ± 0.69) mSv, respectively. Seventeen lumen segments with greater than 50% stenosis had DSA measurement results.

Quantitative analysis

The results of quantitative analysis are shown in Table 1. There were no statistical differences in the CT values measured on the blood vessels and muscles ROI in the four reconstructed image groups, while the image noise (SD value) measured on the blood vessel and muscle ROI were statistically significantly different. The SD value of AV-50 images was the largest, while AV-100 images was the smallest among the four image sets, among which DL-M was slightly smaller than that of AV-50 images, while the SD value of DL-H was slightly higher than that of AV-100 images. The intergroup SNR and CNR comparison of all ROI on each anatomical level showed significant difference, the orders of SNR and CNR metric value were as follows: AV-100 > DL h > DL M > AV-50. The detailed results of the comparison between the two groups are shown in Figure 2.

Table 1.

Quantitative analysis of conventional indexes for four reconstructed images

(n = 46)		AV-50	AV-100	DL-M	DL-H	F	P
Aorta
	HU	462.9 ± 127.4	469.1 ± 111.3	470.5 ± 112.1	470.3 ± 112.3	0.044	0.988
	SD	27.8 ± 6.1	15.3 ± 5.4	23.1 ± 4.9	16.7 ± 4.3	57.112	<0.001
	SNR	17.3 ± 5.8	33.9 ± 12.6	20.9 ± 4.9	29.1 ± 6.9	39.810	<0.001
	CNR	21.4 ± 9.6	42.8 ± 16.2	23.8 ± 8.6	35.0 ± 13.0	41.701	<0.001
Psoas muscles
	HU	59.6 ± 11.4	59.2 ± 11.3	60.6 ± 10.9	60.5 ± 11.0	0.157	0.925
	SD	20.0 ± 4.1	8.6 ± 2.8	17.9 ± 3.1	12.2 ± 2.0	133.015	<0.001
	SNR	3.1 ± 0.8	7.6 ± 2.8	3.5 ± 0.9	5.1 ± 1.1	74.707	<0.001
	CNR	9.3 ± 1.7	23.2 ± 7.0	10.4 ± 2.0	15.3 ± 2.4	118.904	<0.001
Popliteal artery
	HU	482.1 ± 98.1	478.8 ± 95.3	490.8 ± 97.2	491.3 ± 96.9	0.213	0.887
	SD	21.6 ± 9.5	9.8 ± 8.7	20.5 ± 9.0	17.7 ± 8.5	12.178	<0.001
	SNR	22.3 ± 50.1	68.3 ± 51.7	31.3 ± 24.9	40.1 ± 46.5	6.720	<0.001
	CNR	33.3 ± 10.7	85.2 ± 31.5	35.9 ± 13.0	50.7 ± 44.7	7.053	<0.001
Popliteal muscles
	HU	58.1 ± 11.8	57.1 ± 2.0	57.3 ± 10.5	56.8 ± 9.7	0.118	0.949
	SD	13.5 ± 3.3	4.7 ± 2.8	13.1 ± 3.5	9.8 ± 5.4	50.640	<0.001
	SNR	4.6 ± 1.5	15.9 ± 8.1	4.8 ± 1.7	6.7 ± 2.3	70.181	<0.001
	CNR	14.2 ± 3.9	49.8 ± 25.3	14.7 ± 5.0	20.8 ± 6.4	72.258	<0.001
Ankle muscles
	HU	63.3 ± 14.8	62.5 ± 13.5	62.5 ± 13.1	62.0 ± 12.3	0.066	0.978
	SD	12.6 ± 3.7	4.3 ± 2.1	10.8 ± 3.2	7.7 ± 2.3	71.362	<0.001
	SNR	5.5 ± 2.1	18.6 ± 9.9	6.3 ± 2.2	8.9 ± 3.2	56.722	<0.001
	CNR	15.1 ± 4.9	50.7 ± 26.3	17.4 ± 5.5	26.1 ± 1.5	53.153	<0.001

CNR, contrast-to- noise ratio; HU, Hounsfield unit; SD, standard deviation; SNR, signal-to-noise ratio.

Data are the mean ± standard deviation

Figure 2.

Comparison of conventional quantitative indexes of different reconstruction methods (a) shows the comparison of SD values of four groups at different ROI. (b) Shows the comparison of SNR of four groups at different ROI. (c) Shows the comparison of CNR of four groups at different ROI. Ankle-M: Ankle muscles; CNR, contrast-to-noise ratio; Popliteal-A: Popliteal artery; Popliteal-M: Popliteal muscles; ROI, region of interest; SD, standard deviation; SNR, signal-to-noise ratio.

For the sharpness of the vessel edge, it was found that there were statistically significant differences in the overall comparison of ERD and ERS in the four groups. The mean ERD was significantly shorter on the two DLIR images than those of ASIR-V reconstruction, among which the ERD of AV-100 was the longest, while that of DL-M was the shortest. The ERS results of the four reconstructed image groups were exactly the opposite of ERD, and the results were: DL-M > DL h > AV-50 > AV-100. The specific results are shown in Table 2. Figure 3 shows the results of pairwise comparison of ERS and ERD of the four reconstruction methods, and Figure 1 (d) shows the CT attenuation-distance curves of the images of four reconstruction methods of one patient.

Table 2.

Quantitative analysis (evaluation of the profile curves)

(n = 46)	AV-50	AV-100	DL-M	DL-H	F	P
ERD (mm)	1.96 ± 0.39	2.01 ± 0.36	1.79 ± 0.37	1.82 ± 0.38	3.644	0.014
ERS (HU/mm)	195.73 ± 54.15	188.38 ± 49.32	218.91 ± 63.51	214.94 ± 63.49	2.985	0.033

ERD, edge -rise distance; ERS, edge -rise slope.

Data are the mean ± standard deviation

Figure 3.

The results of pairwise comparison of ERS and ERD of the four reconstruction methods “*” indicates there is significant difference between groups (p < 0.05), “* *” indicates there is extremely significant difference between groups (p < 0.01). ERD, edge-rise distance; ERS, edge-rise slope.

Qualitative evaluation

The results of qualitative evaluation of the four reconstruction groups showed statistical differences in granularity and edge blurring. DL-M had the same edge blurring score as the AV-50, but significantly smaller (better) granularity score than AV-50 (0.74 vs 1.76 from Observer 1 and 0.83 vs 1.85 from Observer 2). The granularity score of DL-H was slightly larger than AV-100 (0.11 vs 0.07 from Observer 1 and 0.18 vs 0.15 from Observer 2), but its edge blurring score was significantly smaller than AV-100 (0.09 vs 2.01 from Observer 1 and 0.22 vs 2.22 from Observer 2), and AV-100 images had the smallest granularity score but highest edge blurring score. The edge blurring score of DL-M images was 0 point, with granularity score higher than that of DL-H. The DL-H images had the highest subjective evaluation score. There was substantial interobserver agreement with regard to the overall image quality with K > 0.83. the detailed results of the qualitative evaluation are shown in Table 3. The comparison of the four reconstruction groups is shown in Figures 4 and 5.

Table 3.

Qualitative analysis of image granularity and blurring

	Reader 1						Reader 2
(n = 46)	AV-50	AV-100	DL-M	DL-H	χ2	P	AV-50	AV-100	DL-M	DL-H	χ2	P
Granularity	1.76 ± 0.57	0.07 ± 0.25	0.74 ± 0.53	0.11 ± 0.31	119.473	<0.001	1.85 ± 0.60	0.15 ± 0.36	0.83 ± 0.61	0.18 ± 0.39	106.714	<0.001
Blurring	0	2.01 ± 0.25	0	0.09 ± 0.28	132.750	<0.001	0	2.22 ± 0.73	0	0.22 ± 0.42	126.226	<0.001
Overall	1.76 ± 0.57	2.13 ± 0.78	0.74 ± 0.53	0.19 ± 0.41	120.642	<0.001	1.85 ± 0.60	2.37 ± 0.77	0.83 ± 0.61	0.41 ± 0.62	106.567	<0.001

Figure 4.

Axial images of four reconstruction methods. 67-year-old male. (a, e) Shows the AV-50 images, with a noise score of 2 and a blurring score of 0. (b, f) Shows the AV-100 images, with a noise score of 1 and a blurring score of 2. (c, g) Shows the DL-M images, with a noise score of 0 and a blurring score of 0. (d, h) Shows the DL-H images, with a noise score of 0 and a blurring score of 1.

Figure 5.

MIP images of four reconstruction methods. 67-year-old male. (a) Shows the MIP images of AV-50. (b) Shows the MIP images of AV-100. (c) Shows the MIP images of DL-M. (d) Shows the MIP images of DL-H. MIP, maximum intensity projection.

Evaluation of diagnostic accuracy

The 17 lumen segments with greater than 50% stenosis indicated by DSA measurement were used to calculate the MSE using the DSA results as the reference standard. The DL-H images resulted in the smallest MSE value. The specific results are shown in Table 4. Bland–Altman plots between the CT and DSA results for the four reconstruction types are shown in Figure 6, also indicating that the consistency between DL-H and DSA was the best in the measurement of lower extremity arterial stenosis.

Table 4.

Mean square error for CT stenosis rate measurement against DSA

(n = 17)	AV-50	AV-100	DL-M	DL-H
MSE against DSA	57.86	73.03	31.92	18.62

DSA, digital subtraction angiography; MSE, mean square error; MSE: (SUM(X-DSA)^2)/n.

Figure 6.

Bland–Altman plot for four methods and DSA.We used Bland–Altman plot diagram to evaluate the consistency of the stenosis rates measured by the four methods and DSA. We set 10 as the error limit. (a) Shows that 11.8% (2/17) points are outside the limit; (b) shows that 17.6% (3/17) points are outside the limit; (c) shows that 5.9% (1/17) points are outside the limit; (d) shows that all points are within the limit, and compared with the other three figures, all points are closest to the mean value. DSA, digital subtraction angiography.

Discussion

In this study, we evaluated DLIR in reducing image noise, improving image quality and diagnostic accuracy of detecting lumen stenosis in low-dose CTA of lower extremity arteries in comparison with the ASIR-V algorithm.

Low-dose CTA of lower extremity arteries is a research focus. To compensate for the loss of image quality caused by the reduced signal strength in low radiation dose condition, iterative reconstruction methods are often employed to reduce the image noise. ^11–13 The traditional iterative reconstruction method needs a “forward projection” to reconstruct an estimated value of the raw data, which is further used to correct for the incorrectness in images in the next iteration, and this process is repeated to optimize the final image quality. Studies have shown that, the iterative reconstruction algorithms often generate images with “blotchy” noise texture or “plastic looking” and is more obvious with the increase of iterative intensity. ^3,14–16 It is sometime difficult to balance image noise and spatial resolution in images.

DLIR is a new CT image reconstruction method based on DNNs intent to solve the current challenges of iterative reconstruction algorithms. DLIR is characterized by using standard-dose, high-quality FBP data sets to train DNNs to learn how to distinguish signal from noise, and to effectively suppress noise without negatively affecting the anatomy and pathological structures. Solomon et al compared the noise and spatial resolution of the three reconstruction methods of FBP, ASIR-V and DLIR in a phantom study, ¹⁷ and proved that compared with FBP, DLIR can greatly reduce the image noise while maintaining the noise texture and high contrast spatial resolution (with only 9% reduction in the average spatial resolution frequency), while the spatial resolution frequency of ASIR-V image was significantly reduced (55% lower). The phantom research of Higaki et al ⁸ proved that the noise on the image reconstructed by DLIR was lower than that of the image reconstructed by other methods (FBP, hybrid iterative- and model-based iterative reconstruction), especially in the low radiation doses circumstances, and DLIR also outperformed other methods with respect to task-based detectability in the case of low radiation dose scanning. In addition to the phantom, some scholars have applied DLIR to multiple parts of the human body imaging (such as the chest, abdomen, head, coronary artery, etc.) and have also proved the advantages of DLIR in improving image quality and reducing radiation dose. ^18–24 Park et al ⁷ also compared DLIR and ASIR-V reconstruction methods in imaging lower extremity arteries, and the results showed that DLIR produced the most balanced image in terms of image noise and sharpness among the different reconstruction algorithms. However, their study only focused on the advantages of DLIR in improving image quality but did not evaluate its performance in disease diagnosis.

In our study, we comparatively evaluated the image quality of ASIR-V (at 50 and 100% levels) and DLIR (at medium and high levels) and diagnostic accuracy for stenosis greater than 50% lumen diameter in low-dose CTA of lower extremity arteries. In terms of image quality evaluation, we used conventional quantitative evaluation metrics (CT value, SD value, SNR, CNR) and qualitative image quality scores. By comparing the SD value (reflecting image noise), we found that in all the selected anatomical levels of interest, the AV-100 generated the largest noise reduction compared to AV-50 by 45.17–66.08%, followed by DL-H with 17.98–39.07% reduction. However, in the qualitative evaluation, AV-100 had the highest image blur rating (at 2.01 and 2.22), indicating that although AV-100 could significantly reduce image noise, it also caused the image to be overly too smooth in this application. Since the reconstruction methods of each group did not affect the CT values (p > 0.05), the results of SNR and CNR had negative relationship with the SD value: the larger the SD value, the smaller the SNR and CNR value. In the subjective image quality score, both DL-M and DL-H showed better image quality than ASiR-V (the total scores for edge blurring and granularity were significantly smaller). All these data indicated that while significantly reducing image noise, DLIR maintained the noise texture of the images to avoid “unnatural looking” in images.

The impact of reconstruction algorithms on image spatial resolution was further objectively evaluated in our study using ERD and ERS to reflect the sharpness of blood vessel wall. Suzuki et al ⁹ in their phantom study proposed to use CT attenuation profile through vessels to measure vessel diameter and to obtain the CT attenuation curve of vessel wall, which could be used to reflect the sharpness or blurring of vascular wall. We used the same method to calculate the ERD and ERS of all target vessel segments of images with different reconstruction methods. Our results showed that the two DLIR image groups had sharper vascular walls than ASIR-V images (higher ERS and smaller ERD), while AV-100 had the worst ERD and ERS results. DLIR demonstrated its robustness of balancing noise reduction and edge sharpness preservation over the conventional iterative reconstruction algorithms.

In this study, we also evaluated the diagnostic accuracy for detecting stenosis of lower limb arteries greater than 50% lumen diameter in some patients and compared the MSE of the area stenosis rate measured on the four different reconstructed images against DSA results. The results showed that the degrees of stenosis measured on DL-H images were the closest to DSA, while the AV-100 had the largest deviation.

Our study had some limitations. First, our study only compared ASIR-V and DLIR algorithms and did not choose FBP as the benchmark for comparative research. We also did not compare DLIR with other iterative reconstruction algorithms. Second, in terms of the diagnosis and evaluation of stenosis, the number of suitable cases was too small that might produce some bias, and more appropriate cases need to be included for further research. Third, all our evaluation results were based on low-radiation dose scanning, and the relevant comparison in conventional dose scanning conditions requires further investigation.

Conclusion

In conclusion, comprehensive subjective evaluation and objective evaluation show that DLIR images perform better with lower image noise and sharper vessel wall than the standard ASIR-V at 50% blending level in low-dose lower extremity CTA, and DL-H provides the best overall image quality and highest accuracy in diagnosing artery stenoses.