Application of deep learning image reconstruction in low-dose chest CT scan

Abstract

Objective:

Deep learning image reconstruction (DLIR) is a new reconstruction method for maintaining image quality at reduced radiation dose. The purpose of this study was to compare image quality of reduced-dose DLIR images with the standard-dose adaptive statistical iterative reconstruction (ASIR-V) images in chest CT.

Methods:

Our prospective study included 48 adult patients (30 women and 18 men, mean age ±SD, 49.8 ± 14.3 years) who underwent both the standard-dose CT (SDCT) and low-dose CT (LDCT) on a GE Revolution CT scanner. All patients gave written informed consent. All scans were reconstructed with ASIR-V40%. Additionally, LDCT scans were reconstructed with DLIR with high-setting (DLIR-H) and medium-setting (DLIR-M). Image noise and contrast-noise-ratio (CNR) of thoracic aorta with different reconstruction modes were measured and compared.

Results:

LDCT reduced radiation dose by 96% compared with SDCT (CTDIvol: 0.54mGy vs 12.46mGy). In LDCT, DLIR significantly reduced image noise compared with the state-of-the-art ASIR-V40% with DLIR-H provided the lowest image noise and highest image quality score. In addition, the image noise, CNR of aorta and overall image quality of the low-dose DLIR-H images did not have significant difference compared with the SDCT ASIR-V40% images (all p > 0.05).

Conclusion:

DLIR significantly reduces image noise in LDCT chest scans and provides similar image quality as the SDCT ASIR-V images at 4% of the radiation dose.

Advances in knowledge:

DLIR uses high-quality FBP data to train deep neural networks to learn how to distinguish between signal and noise, and effectively suppresses noise without affecting anatomical and pathological structures. It opens a new era of CT image reconstruction. DLIR significantly reduces image noise and improves image quality compared with ASIR-V40% under same radiation dose condition. DLIR-H achieves similar image quality at 4% radiation dose as ASIR-V40% at standard-dose level in non-contrast chest CT.

Introduction

In recent years, lung cancer has become the number one cause of death in several countries. The American National Lung Screening Trial (NLST) has demonstrated that compared with screening using chest X-rays, using low-dose CT (LDCT) has led to 20% reduction in lung cancer mortality by screening heavy smokers. ^1,2 CT has replaced conventional chest radiographs as the preferred method for chest examination. Thus, disease can be controlled or treated in its asymptomatic state. ³ However, compared to chest X-ray, patients may receive 10–100 times of radiation dose when a standard-dose chest CT scan is used. The potential risk of radiation-induced malignancy related to the burgeoning use of CT and radiation exposure has attracted a lot of attention. ² In CT examination, the radiation dose delivered to the patients is a public health concern. ⁴ The radiation related risk may be increasing due to the increase of CT examination. Low-dose chest CT is considered as a screening method for early detection of lung cancer in the population at risk. ⁵ So low-dose chest CT scan has been used more frequently than the standard-dose chest CT in clinical application in the early lung cancer screening, ⁶ and reducing the radiation dose while maintaining or improving image quality is a goal many people have been pursuing. Filtered back projection (FBP) was the standard reconstruction for CT which has the fastest image reconstruction time. However, when images are reconstructed using conventional FBP under low-dose scan conditions, high image noise and artifacts cannot be ignored. ^7,8 Iterative reconstruction techniques have been introduced that can significantly reduce the image noise and provide more possibilities for reducing the radiation dose, compared with the FBP reconstruction. ⁶ The adaptive statistical iterative reconstruction (ASIR-V) is a new generation iterative reconstruction algorithm that has advantages of further improving image quality and/or reducing radiation dose compared with its predecessor. ⁸ However, in general, iterative reconstructions at high levels may cause a plastic-looking, blotchy, and unnatural image appearance that will eventually reduce the image quality and affect clinicians’ diagnosis of diseases, limiting the ability for deep radiation dose reduction.

Recently, a deep learning image reconstruction (DLIR) algorithm (TrueFidelity^TM, GE Healthcare Waukesha, WI) has been introduced to address some of the unsolved difficult scientific and technical problems of iterative reconstruction algorithms. DLIR is a CT image reconstruction method applied with a deep convolutional neural network (DCNN) to improve image quality. ⁹ DLIR uses highly selected, essentially artifact-free FBP image sets of both phantoms and patients to train the software and has been shown to improve image quality or maintain image quality under lower radiation doses in the abdominal and coronary CT applications. ^2,10 The purpose of our study was to evaluate image quality in terms of image noise, contrast-to-noise ratio (CNR) of DLIR chest CT images under an extremely low-dose scan condition and compare with those of ASIR-V40% images under both standard- and low-dose conditions.

Methods and materials

Study population

This was a prospective study approved by the internal review board of our hospital and all patients gave written informed consent for participating the study. From May 10 to September 28, 2020, 49 consecutive patients (18 men, 31 women; with mean age ± standard deviation of 49.8 ± 14.3 years; body mass index (BMI) of 18.5–24.0 kg/m²) underwent an extremely LDCT scan of the chest after a clinically indicated standard-dose CT (SDCT). One patient was later excluded from analysis due to motion artifacts. Thus, 48 patients (18 men, 30 women) were finally included in this study.

Scan technology and image acquisition

All patients were scanned on a 256-slice CT scanner (Revolution CT, GE Healthcare) while in supine position with arms raised overhead to prevent artifacts. All patients were instructed to avoid any voluntary motion and to carefully follow the breath-hold instructions. The standard- and low-dose scan protocols were used in an inspiratory breath-hold cycle to ensure that the lesions were in the same position in both scans. The scan parameters were as follows: (1), standard-dose scanning: voltage, 120 kV; and automatic tube current; gantry rotation time, 0.5 s; helical pitch, 0.992:1. (2), low dose scanning: voltage, 80 kV; tube current, 50mA; gantry rotation time, 0.5 s; helical pitch, 0.992:1. Both LDCT and SDCT images were reconstructed at a slice thickness of 1.25 mm and with ASIR-V at a strength level of 40% (ASIR-V40%). In addition, the LDCT scan data sets were reconstructed with DLIR at the medium (DLIR-M) and high (DLIR-H) levels.

To assess the radiation dose, the volume CT dose index (CTDI_vol) and dose–length product (DLP) were recorded for the SDCT and LDCT imaging series. The estimated effective dose (ED) was calculated as DLP multiplied by a k-factor of 0.014 mSv·mGy^–1·cm^–1 for the chest.

Objective image analysis

All reconstructed (ASIR-V40%@LDCT, ASIR-V40%@SDCT, DLIR-M@LDCT, DLIR-H@LDCT) images were transmitted to a GE AW 4.7 workstation for data measurement and image analysis. Images were reviewed in both a lung setting (window level, –600 HU; window width, 1500 HU) and mediastinal setting (window level, 40 HU; window width, 350 HU). One radiologist with 3 years of working experience in medical imaging performed an objective image analysis on the axial images. The reconstructions were linked so that identical region of interest (ROI) could be drawn in the same location on each reconstruction. Three ROIs with area of 150 mm² were drawn in the aorta, paraspinal musculature, and subcutaneous fat. For each reconstruction, the contrast-to-noise ratio (CNR) relative to muscle was calculated for the aorta as (ROIi – ROIm) / SD, where ROIi is the mean attenuation for the anatomy of interest (aorta), ROIm is the mean attenuation of paraspinal muscles, and SD is the mean image noise based on the measurement for subcutaneous fat, calculated as the mean SD of attenuation in HUs. These measurements were performed in all four reconstruction modes. ²

Subjective image analysis

Two other radiologists with more than 10 years of experience in medical imaging, separately performed the lesion identification and objective image analysis of the reconstructed images. The patients’ information and the image reconstruction modes were hidden. The radiologists were blinded to the patients’ data and image reconstruction techniques. A 5-point scoring system for subjective evaluation of image quality, ⁶ including aspects of morphological display, visibility for surrounding lung tissue and diagnostic confidence for lung lesions (including solid nodules and ground-glass nodules) was used (Table 1).

Table 1.

Subjective score criteria for image quality evaluation

Grading score	Qualitative image analysis
	Morphological display of all nodules	Visibility for surrounding lung tissue	Artifacts and diagnostic confidence
1	Very poor display, unclear edge	Unacceptable visibility, cannot distinguish small structures	Severe artifacts, insufficient confidence
2	Poor display, fuzzy edge	Small structures are not displayed very well, seriously impact diagnosis	Substantial artifacts, insufficient confidence
3	Moderate display, not very clear edge	Small structures can be displayed, and enough for diagnosis	Moderate artifacts, low confidence but diagnosis possible
4	Better display, still clear edge	Small structures can be clearly displayed with good contrast	Minor artifacts, good diagnostic confidence
5	Excellent display, clear edge	Small structures can be clearly displayed with excellent contrast	No artifacts, excellent diagnostic confidence

Statistical analysis

Data were recorded in Excel (Microsoft Office 2016) and analyzed with SPSS statistical software (v. 22.0, IBM SPSS Statistics). The objective data were expressed as mean ± SD. Radiation dose between LDCT and SDCT was compared using the Student’s t-test. The differences among the CT images reconstructed with ASIR-V40% (with LDCT and SDCT data), DLIR-M, DLIR-H (with LDCT data) were evaluated. The one-way ANOVA with Bonferroni Correction was used to compare the quantitative CT measurements and the Kruskal–Wallis Wilcoxon rank sum test and Dunnett’ t-test was used to compare image quality across the different dose levels and different reconstruction techniques for chest CT. For the subjective analysis, we calculated the interobserver agreement using the κ statistic to evaluate the agreement between the two readers. A p-value of less than 0.05 was considered statistically significant.

Results

Basic information of patients and radiation dose

A total of 48 patients (18 men, 30 women) were finally included in this study. A total of 97 solid nodules and 30 ground-glass nodules were identified in the study in SDCT. There was no significant difference in the detection rate of nodules between the SDCT, LDCT, DLIR-M and DLIR-H. The difference was in the appearance of the nodules.

From the images, the average X-ray tube current of SDCT was 413.78 ± 4.81 mA. As for the radiation dosage, the mean CTDIvol, DLP, ED were 12.46 ± 1.16 mGy, 447.32 ± 34.51mGy*cm, 6.26 ± 0.48 mSv in SDCT and 0.54 ± 0.00 mGy, 19.44 ± 1.37mGy*cm, 0.27 ± 0.20 mSv in LDCT, respectively with about 96% dose reduction in LDCT (all p < 0.001) (Table 2).

Table 2.

Comparison of radiation dose between the two scanning modes

	CTDIvol (mGy）	DLP (mGy.cm)	ED (mSv)
Standard-dose scan	12.46±1.16	447.32±34.51	6.26±0.48
Low-dose scan	0.54±0.00	19.44±1.37	0.27±0.20
p	<0.001	<0.001	<0.001

CTDIvol, volume CT dose index; DLP, dose–length product; ED, effective dose.

Objective analysis

The objective image analysis results are presented in Table 3. The image signals (CT number) were conformed to normal distribution and did not have significantly difference across different reconstructions (p = 0.2). But the other parameters, including image noise (SD), and CNR, differed significantly. For image noise, the SD value did not have significantly difference between ASIR-V40%@SDCT and DLIR-H@LDCT (p = 1.000), while the differences between any other reconstruction pairs were all statistically significant (DLIR-M@LDCT vs ASIR-V40%@LDCT, p = 0.006 and ASIR-V40%@SDCT vs DLIR-M@LDCT, ASIR-V40%@SDCT vs ASIR-V40%@LDCT, DLIR-H@LDCT vs DLIR-M@LDCT, DLIR-H@LDCT vs ASIR-V40%@LDCT, all p < 0.001 (Figure 1)). The CNR for the aorta did not have statistically significant difference between ASIR-V40%@SDCT and DLIR-H@LDCT, (p = 1.000), DLIR-M@LDCT and ASIR-V40%@LDCT, (p = 0.625), and ASIR-V40%@SDCT and DLIR-M@LDCT, (p = 0.163), and DLIR-H@LDCT and DLIR-M@LDCT, (p = 0.181); while there was statistically significant difference between ASIR-V40%@SDCT and ASIR-V40%@LDCT and between DLIR-H@LDCT and ASIR-V40%@LDCT (all p < 0.05).

Table 3.

Comparison of quantitative measurements among ASIR-V and DLIR under different radiation doses

Variables	SDCT	LDCT	DLIR-M	DLIR-H		p
					P	SDCT vs DLIR-H	SDCT vs DLIR-M	SDCT vs LDCT	DLIR-H vs DLIR-M	DLIR-H vs LDCT	DLIR-M vs LDCT
Image noise	12.4 ± 2.1	28.6 ± 3.9	20.4 ± 2.6	12.5 ± 7.2	<0.001	1.000	<0.001	<0.001	<0.001	<0.001	0.006
CNR aorta	0.9 ± 0.6	0.4 ± 0.3	0.6 ± 0.5	0.9 ± 0.7	<0.001	1.000	0.163	0.004	0.181	0.007	0.625
Image signal	46.9 ± 6.6	49.0 ± 9.5	50.2 ± 8.9	50.4 ± 8.0	0.2	0.098	0.124	0.325	0.905	0.497	0.576

ASIR, adaptive statistical iterative reconstruction; DLIR, deep learning image reconstruction; DLIR-H: DLIR-H with low-dose; DLIR-M: DLIR-M with low-dose; LDCT: ASIR-V40% with low-dose;LDCT, low-dose CT; SDCT, standard-dose CT; SDCT: ASIR-V40% with standard-dose.

Figure 1.

Comparison of chest CT scan in axial soft tissue window images of mediastinum in 43-year-old male (A), ASIR-V40% at SDCT; (B), ASIR-V40% at LDCT; (C), DLIR-M at LDCT; and (D), DLIR-H at LDCT. In different reconstructions, the image attenuation values (CT numbers) did not have statistically significant difference. For image noise, the SD value did not have significantly difference between SDCT and DLIR-H (p = 1.000), while there were statistically significant differences between any other reconstruction pairs: DLIR-M vs LDCT, p = 0.006, and SDCT vs DLIR-M, SDCT vs LDCT, DLIR-H vs DL-M, DLIR-H vs LDCT, all p < 0.001. ASIR, adaptive statistical iterative reconstruction; DLIR, deep learning image reconstruction; LDCT, low-dose CT; SDCT, standard-dose CT; SD, standard deviation.

Subjective analysis

The subjective analysis is summarized in Tables 4 and 5 and Figure 2 . The results showed that, There was no significant difference in the detection rate of nodules among the different reconstructions. The difference was in the appearance of the nodules.The DLIR images had a better quality than the ASIR-V40% at LDCT. Moreover, compared with SDCT, DLIR-H images at LDCT had similar image quality for evaluating the nodules in terms of morphological display of nodules, visibility for surrounding lung tissue, artifacts, and diagnostic confidence (Tables 4 and 5). There was substantial agreement between the two readers (Kappa>0.7).

Table 4.

Subjective image analysis results with ASIR-V and DLIR under different radiation doses

Variables	SDCT	LDCT	DLIR-M	DLIR-H	P
					P	SDCT vs DLIR-H	SDCT vs DLIR-M	SDCT vs LDCT	DLIR-H vs DL-M	DLIR-H vs LDCT	DLIR-M vs LDCT
Solid nodule	4.20 ± 0.40	2.72 ± 0.62	3.31 ± 0.73	4.08 ± 0.49	<0.001	0.43	<0.001	<0.001	<0.001	<0.001	<0.001
Ground-glass nodule	3.70 ± 0.47	2.10 ± 0.31	2.33 ± 0.48	3.47 ± 0.82	<0.001	0.34	<0.001	<0.001	<0.001	<0.001	0.16
Mediastinal tissue	4.12 ± 0.33	2.04 ± 0.20	3.02 ± 0.44	3.85 ± 0.41	<0.001	<0.005	<0.001	<0.001	<0.001	<0.001	<0.001

ASIR, adaptive statistical iterative reconstruction; DLIR, deep learning image reconstruction; LDCT, low-dose CT; SD, standard deviation; SDCT, standard-dose CT.

Data are presented as mean ± SD. p < 0.05; SDCT: ASIR-V40% with standard-dose; LDCT: ASIR-V40% with low-dose; DLIR-M: DLIR-M with low-dose; DLIR-H: DLIR-H with low-dose.

Table 5.

Subjective scoring for the whole image quality of all pulmonary nodules among ASIR-V and DLIR under different radiation doses

Variables	SDCT	LDCT	DLIR-M	DLIR-H	P
					P	SDCT vs DLIR-H	SDCT vs DLIR-M	SDCT vs LDCT	DLIR-H vs DL-M	DLIR-H vs LDCT	DLIR-M vs LDCT
Morphological display of nodules	4.37 ± 0.49	2.71 ± 0.51	3.37 ± 0.49	4.32 ± 0.52	<0.001	0.66	<0.001	<0.001	<0.001	<0.001	<0.001
Visibility for surrounding lung tissue	4.41 ± 0.50	2.44 ± 0.55	3.24 ± 0.58	4.29 ± 0.56	<0.001	0.32	<0.001	<0.001	<0.001	<0.001	<0.001
Artifacts and diagnostic confidence	4.36 ± 0.49	2.46 ± 0.55	3.34 ± 0.62	4.27 ± 0.50	<0.001	0.41	<0.001	<0.001	<0.001	<0.001	<0.001

ASIR, adaptive statistical iterative reconstruction; DLIR, deep learning image reconstruction; DLIR-H: DLIR-H with low-dose; DLIR-M: DLIR-M with low-dose; LDCT: ASIR-V40% with low-dose;LDCT, low-dose CT; SD, standard deviation; SDCT, standard-dose CT; SDCT: ASIR-V40% with standard-dose.

Figure 2.

Comparison of chest CT scan in axial soft tissue window images of lung in 42-year-old male. Images were (A), ASIR-V40% at SDCT; (B), ASIR-V40% at LDCT; (C), DLIR-M at LDCT; and (D), DLIR-H at LDCT. There was no significant difference in the detection rate of nodules among the different reconstructions. The difference was in the appearance of the nodules. ASIR, adaptive statistical iterative reconstruction; DLIR, deep learning image reconstruction; LDCT, low-dose CT; SDCT, standard-dose CT; SD, standard deviation.

Discussion

During the history of CT development, people have made unremitting efforts to reduce radiation dose while maintaining the image quality and diagnostic accuracy. FBP was the standard reconstruction for CT. However, with the decrease of radiation dose, image quality was greatly affected, prominent noise and artifacts occurred. ¹¹ Then, several methods of maintaining image quality while reducing radiation dose were introduced in succession, such as IR, model-based IR (MBIR), ASIR and ASIR–V (GE Healthcare). ¹² IR techniques have been introduced to reduce image noise or maintain good CT image quality on reduced-dose CT scans. ASIR (GE Healthcare) was the first commercially available IR algorithm. ¹³ One can choose the percentages of blending IR with FBP to obtain the desired balance between noise reduction, spatial resolution and image appearance for clinical application, and is a major advance in the development of reconstruction technology. ¹¹ MBIR is a more advanced iterative algorithm than ASIR, using both backward and forward projections. MBIR can reduce image noise more effectively than ASIR, through many complex models, such as system noise model, object model, physics model. Recent studies have shown that MBIR allows significant reduction of radiation dose without affecting image quality and has the potential to further increase the detection rate of some subtle lesions at the expense of longer reconstruction time. ¹⁴ ASIR-V is the vendor’s third-generation IR algorithm, and replaces its first-generation IR algorithm, ASIR. ASIR-V contains improved noise and object modeling compared with ASIR. But the image noise reduction potential of ASIR-V is lower than that of the MBIR. However, compared with MBIR, the reconstruction time of ASIR-V is substantially reduced, which is one of the major limitations for clinical use of MBIR. ¹⁵ Deep learning is a subset of machine learning in artificial intelligence. In general, deep learning consists of massive multilayer networks of artificial neurons. And, the deep convolutional neural networks (DCNN) method, is commonly used in image recognition. ¹⁶ The DCNN is trained with virtual low- and high-quality images, the former is obtained with LDCT and the latter with SDCT. Compared with conventional machine learning methods, a distinctive feature of deep learning is that it can generate appropriate models for tasks directly from the raw data, removing the need for human-led feature extraction, and images reconstructed with DLIR have the property of reduced image noise without blurring. ^17,18 The DLIR technique assessed in our study represents a major advancement in the pursuit of CT radiation dose optimization. Studies have shown that compared with both conventional FBP and IR techniques, DLIR-based images delivered better qualitative and quantitative image quality while enabling superior lesion detection ability on chest LDCT ¹⁹ and that deep learning approaches offer the exciting potential to more complex image analysis, detect subtle holistic imaging findings and unify methodologies for image evaluation. ²⁰

In our study, we assessed the use of 80 kV and 50 mA for a low-dose chest CT, which reduced radiation dose by 96% compared with the SDCT. We evaluated the image quality of DLIR chest CT images under such a low-dose scan condition and compared with that of ASIR-V40% images under both the standard- and low-dose conditions. Our study demonstrated that, the attenuation values (CT numbers) in images had no significant difference among different reconstructions (SDCT, LDCT, DLIR-M and DLIR-H); However, under the same low-dose condition, DLIR significantly reduced image noise, resulting in higher CNR compared with ASIR-V40%; In addition, DLIR at the 4% radiation dose level provided similar image quality as ASIR-V40% at the standard dose level.

Our study also indicated that although all lesions could be displayed in all the reconstructions studied, the overall image quality, lesion diagnostic confidence, artifacts, image noise and texture and the details of the lesion among the different image reconstructions varied greatly. DLIR showed improved image quality compared to ASIR-V40% in low-dose chest CT scans. For the images at LDCT, the lesion diagnostic confidence was significantly higher with DLIR-H than with ASIR-V40% or DLIR-M. There was no significant difference in lesion diagnostic confidence between ASIR-V40% at SDCT and DLIR-H at LDCT for solid nodules (p = 0.43) and for ground-glass nodules (p = 0.34). While there were statistically significant differences between any other reconstruction pairs: DLIR-M@LDCT vs ASIR-V40%@LDCT; ASIR-V40%@SDCT vs DLIR-M@LDCT; ASIR-V40%@SDCT vs ASIR-V40%@LDCT; DLIR-H@LDCT vs DLIR-M@LDCT; and DLIR-H@LDCT vs ASIR-V40%@LDCT, all p < 0.05.

There were limitations in our study. We used a lower tube voltage (at 80 kVp) in LDCT than the 120 kVp in SDCT to dramatically reduce radiation dose. In the future, we plan to investigate the dose saving and image quality improvement potential using the same tube voltage to reduce variables. Another limitation of our study pertains to the fact that we only had small number of patients and we only included non-contrast chest CT examinations. Studies with more patients and contrast-enhanced CT scans need to be carried out in the future to generalize our conclusion.

Conclusions

In summary, DLIR significantly reduces image noise in low-dose chest CT scans and DLIR-H provides similar image quality as the SDCT ASIR-V40% images with only 4% of the radiation dose.