Application of deep learning reconstruction at prone position chest scanning of early interstitial lung disease

Abstract

Aim

Timely intervention of interstitial lung disease (ILD) was promising for attenuating the lung function decline and improving clinical outcomes. The prone position HRCT is essential for early diagnosis of ILD, but limited by its high radiation exposure. This study was aimed to explore whether deep learning reconstruction (DLR) could keep the image quality and reduce the radiation dose compared with hybrid iterative reconstruction (HIR) in prone position scanning for patients of early-stage ILD.

Methods

This study prospectively enrolled 21 patients with early-stage ILD. All patients underwent high-resolution CT (HRCT) and low-dose CT (LDCT) scans. HRCT images were reconstructed with HIR using standard settings, and LDCT images were reconstructed with DLR (lung/bone kernel) in a mild, standard, or strong setting. Overall image quality, image noise, streak artifacts, and visualization of normal and abnormal ILD features were analysed.

Results

The effective dose of LDCT was 1.22 ± 0.09 mSv, 63.7% less than the HRCT dose. The objective noise of the LDCT DLR images was 35.9–112.6% that of the HRCT HIR images. The LDCT DLR was comparable to the HRCT HIR in terms of overall image quality. LDCT DLR (bone, strong) visualization of bronchiectasis and/or bronchiolectasis was significantly weaker than that of HRCT HIR (p = 0.046). The LDCT DLR (all settings) did not significantly differ from the HRCT HIR in the evaluation of other abnormal features, including ground glass opacities (GGOs), architectural distortion, reticulation and honeycombing.

Conclusion

With 63.7% reduction of radiation dose, the overall image quality of LDCT DLR was comparable to HRCT HIR in prone scanning for early ILD patients. This study supported that DLR was promising for maintaining image quality under a lower radiation dose in prone scanning, and it offered valuable insights for the selection of images reconstruction algorithms for the diagnosis and follow-up of early ILD.

Introduction

Interstitial lung disease (ILD) is a group of diseases caused by diffuse lung parenchymal disease, characterized by chronic inflammation of parabronchial tissues, alveolar spaces, and adjacent supporting structures, and interstitial fibrosis [1, 2]. ILD can lead to lung remodeling, progressive lung function deterioration, a limited response to treatment, and even respiratory failure [3]. Timely intervention, including drugs and nonmedical treatment during the early stage of ILD, is promising for attenuating lung function decline and improving clinical outcomes [4]. To evaluate the suspected and known ILD, high-resolution CT (HRCT) was widely used, and prone position scanning was considered as a reliable protocol to differentiate pathological features from gravitation-dependent density in early ILD patients [5–8].  A survey performed by European Society of Thoracic Imaging reported that 59% of the responders would perform additional prone position scanning in ILD evaluation [9]. However, repetitive scanning during the follow up of ILD can cause accumulative radiation exposure, which is correlated with a greater risk of cancer and other diseases [10]. Therefore, clinicians must weigh the benefits and risks of duplicated HRCT scans.

With the wide use of supercomputing power, a newly developed reconstruction method called deep learning reconstruction (DLR) has emerged. This technique is based on a deep convolutional neural network (DCCN), which was trained with noise-contaminated and noise-free training pairs to allow the network to extract signals from low quality hybrid IR images to have similar image quality as high-dose model-based IR images. During the training process which can be described as solving optimization problem, DCNN incorporates advanced physics models (e.g., statistical model, scanner model, optics model and cone angle model) through learning from the target images. Therefore, DLR images could reduce the image noise and artifact even if in low dose condition [11]. Multiple phantom and clinical studies have shown that DLRs maintain or improve image quality at lower doses or at a given dose [11–14]. However, studies applying the DLR in early-stage ILD patients are limited. Therefore, this study explored whether DLR could maintain image quality and reduce radiation dosage for prone position scanning in early ILD patients.

Methods

Study design

This prospective study had been approved by our Institutional Review Board, and written informed consent was obtained from each participant (Ethics approval reference number: HS-2175). The study was conducted in accordance with the Declaration of Helsinki. Patients were admitted to the outpatient clinic between December 2020 and September 2021. The inclusion criteria were as follows [15]: (1) typical findings of ILD in previous chest CT images; (2) no respiratory symptoms or pulmonary function test (PFT) impairment, and a diagnosis of early interstitial lung disease by a pulmonologist within the last year; (3) no other underlying lung disease (such as COPD, lung infection uncured, malignancies); and (4) need for a prone HRCT scan during a routine return visit. The exclusion criteria included: (1) inability to hold breath to finish a standard CT scan, (2) a history of lung surgery. (3) no interstitial lung abnormality (ILA) on the prone CT imaging. Finally, 21 ILD patients, including eight males and thirteen females, were enrolled.

Scan protocols and reconstruction methods

All patients were scanned with a 320-row detector CT (Aquilion ONE Genesis, Canon Medical System). Every patient underwent two consecutive unenhanced chest CT scans at 20-second intervals with randomized order. The HRCT scan was performed using a 120 kVp tube voltage, automatic tube current, and standard deviation of 8.8 HU. The LDCT scan was performed using a 120 kVp tube voltage and 30 mAs tube current. Other parameters were kept unchanged, including rotation time of 0.5 s/circle, detector number 0.5 mm × 80, and pitch standard (PF 0.813/HP 65.0). Scanning was performed from the diaphragm to the pulmonary apex in the craniocaudal direction and prone position. Patients were required to hold their breath after full inspiration during the scan.

Image reconstruction

HRCT images were reconstructed with standard reconstruction method in clinical diagnosis, called HIR (Adaptive Iterative Dose Reduction 3-Dimensional, AIDR3D) with a standard lung kernel, and LDCT images were reconstructed with DLR (Advanced Intelligence Clear-IQ Engine, AiCE) with lung/bone kernels at mild, standard, or strong levels. All images were reconstructed in the axial plane, and the slice thickness and increment were both 1 mm.

Objective image quality assessment

Objective image quality was assessed by measuring the mean CT value and standard deviation of spherical volume regions of interest (VOI). As the previous study described [16], the VOI within the tracheal lumen were segmented, and mean CT value and standard deviation of VOI were recorded. The standard deviation (SD) was defined as the image noise.

Subjective image quality assessment

The subjective assessment included the overall image quality, subjective image noise, streak artifacts, and normal and abnormal lung structures. All patient information and scanning parameters were hidden and renumbered before the assessment. Two trained radiologists with three and five years of experience in ILD independently evaluated these images at a preselected lung window setting (window width 1200 HU, window level − 600 HU). No changes in these settings were allowed during the assessment to preclude any influence on image quality. When there was any uncertainty, one radiologist with over 40 years of experience would be turned to for help.

The overall image quality was scored using a five-point scale for the whole lung and each lung lobe:

1 score: Excellent (no artifacts or excessive noise, clear anatomical structures);

2 score: Good (minor artifacts, localized and unifocal loss of sharpness);

3 score: Fair (moderate artifacts without loss of diagnostic information, localized plurifocal loss of sharpness);

4 score: Poor (major artifacts with minor loss of diagnostic information);

5 score: Very poor (major artifacts with significant loss of diagnostic information).

The image noise and streak artifacts were scored on a three-point scale:

1 score: Unapparent or minimal;

2 score: Moderate but did not alter the identification of normal and/or abnormal structures;

3 score: Severe and altered the identification of normal and/or abnormal structures.

Normal lung features, including fissures, proximal bronchi and adjacent pulmonary vessels (above subsegmental level), peripheral bronchi and adjacent pulmonary vessels, and subpleural vessels, were evaluated.

For the abnormal lung features to be chosen, as interstitial lung abnormality (ILA) was considered as possible signs of early ILD based on systematic evaluation of large cohorts, which was defined as incidental finding of nondependent abnormalities affecting more than 5% of any lung zone at CT, including ground-glass or reticular abnormalities, lung distortion, traction bronchiectasis or bronchiolectasis, honeycombing, and nonemphysematous cysts, however, no consensus on the report of ILA had been reached until now [4, 17] so we only selected several representative features in this study to evaluate, including ground glass opacity (GGO), reticulation, bronchiectasis and bronchiolectasis, honeycombing, and architectural distortion, and these abnormal lung features were evaluated at the lobar level. Both normal and abnormal structures were scored on a five-point scale:

0 score: No above abnormal lung features (only applied to abnormal structures evaluation).

1 score: Excellent image quality with sharp demarcation of structures;

2 score: Slight blurring of the structures, with unrestricted image evaluation;

3 score: Moderate blurring of the interface structures, with slightly restricted assessment;

4 score: Severe blurring and poorly defined structures, with uncertainty about the evaluation;

5 score: Severe reduced image quality, making reliable interpretation impossible [18–21].

Radiation dosage

The dose-length product (DLP) and volume CT dose index (CTDI_vol) were extracted from patient protocols. The lateral (LAT) and anteroposterior (AP) diameters were measured at the mid-liver level of axial CT images with digital calipers, and the square root of the product of LAT and AP was defined as the effective diameter. The effective radiation dose (ED) was the product of DLP and a chest-specific conversion coefficient (0.014 mSv/mGy·cm). [22, 23] The size-specific dose estimates (SSDE) were the product of CTDI_vol and a size-specific conversion factor f_size, obtained based on the effective diameter by checking the AAPM Report 204 [24].

Statistical analysis

Continuous variables were calculated as the mean ± standard deviation (SD), and categorical variables were summarized as numbers and frequencies. The interobserver agreement of the two radiologists was assessed using Cohen’s κ-test. The κ value was classified as poor (0–0.20), fair (0.21–0.40), moderate (0.41–0.60), substantial (0.61–0.80), or excellent (0.81–1.00) [25]. The image quality and radiation dosage of the reconstructed HRCT and LDCT images were compared. All analyses were performed using paired sample t-tests or the Wilcoxon signed-rank test according to the results of the Shapiro-Wilk test. The data were analysed in IBM SPSS 20.0. The two-tailed statistical significance threshold was set at p < 0.05.

Results

Twenty-one patients with early-stage ILD were enrolled, with a mean age of 62.2 ± 7.2 years (range: 45–78 years) and a mean BMI of 25.69 ± 3.05 kg/m². Seven patients were diagnosed with connective tissue disease-related (CTD-related) ILD, and the other patients were of unknown origin.

Radiation dosage

Compared with HRCT group, the LDCT scans had significantly lower CTDI_vol, DLP, ED and SSDE (p < 0.001), with a 63.7% radiation dose reduction. (Table 1)

Table 1.

Radiation dose Estimation of HRCT and LDCT images

	HRCT	LDCT	p
CTDIvol (mGy)	5.65 ± 1.35	2.00	< 0.001
DLP (mGy·cm)	239.86 ± 70.80	87.02 ± 6.61	< 0.001
ED (mSv)	3.36 ± 0.84	1.22 ± 0.09	< 0.001
SSDE (mGy)	7.57 ± 1.33	2.72 ± 0.20	< 0.001

Data are presented as the mean ± SD

CTDI_vol, volume CT dose index; DLP, dose-length product; ED, effective radiation dose; SSDE, size-specific dose estimates

Objective image quality assessment

The LDCT DLR images (apart from the mild and standard lung algorithm) had significantly less image noise than the HRCT HIR images (p < 0.001 or p < 0.01, Table 2).

Table 2.

Objective image quality assessment of HRCT HIR and LDCT DLR

	CT (HU)	p*	SD (HU)	p*
HRCT-AIDR3D	-977.295.01		44.178.57
LDCT-AiCE (lung, mild)	-989.178.56	0.008	49.756.98	0.084
LDCT-AiCE (lung, standard)	-989.408.81	0.008	41.027.09	0.225
LDCT-AiCE (lung, strong)	-987.717.55	0.015	30.683.58	0.005
LDCT-AiCE (bone, mild)	-986.689.33	0.065	29.336.89	0.003
LDCT-AiCE (bone, standard)	-984.2110.72	0.087	18.608.84	< 0.001
LDCT-AiCE (bone, strong)	-981.5513.15	0.107	15.8410.62	< 0.001

Data are presented as the mean ± SD

HIR, hybrid iterative reconstruction; DLR, deep learning reconstruction; SNR, signal-to-noise ratio

* compared with HRCT AIDR3D

Subjective image quality assessment

The interobserver agreement was substantial (κ value = 0.73). There was no significant difference in the overall image quality of whole lung or lobe pairs between HRCT HIR and LDCT DLR (Table 3; Fig. 1). Subjective image noise in the LDCT DLR group (lung/bone, standard/strong) was significantly lower than that in the HRCT HIR group(p < 0.001). Streak artifacts on LDCT DLR were not significantly different from those on HRCT HIR (Table 3).

Table 3.

Whole lung overall image quality assessment for HRCT HIR and LDCT DLR

Scores	HRCT HIR	LDCT DLR (lung, mild)	p	LDCT DLR (lung, standard)	p	LDCT DLR (lung, strong)	p	LDCT DLR (bone, mild)	p	LDCT DLR (bone, standard)	p	LDCT DLR (bone, strong)	p
Overall image quality
1	1(4.8)	1(4.8)	0.317	1(4.8)	0.317	1(4.8)	0.317	1(4.8)	0.317	1(4.8)	> 0.999	1(4.8)	> 0.999
2	12(57.1)	12(57.1)		12(57.1)		12(57.1)		12(57.1)		11(52.4)		11(52.4)
3	7(33.3)	8(38.1)		8(38.1)		8(38.1)		8(38.1)		9(42.9)		9(42.9)
4	1(4.8)	0		0		0		0		0		0
5	0	0		0		0		0		0		0
Image noise
1	0	1(4.8)	0.317	12(57.1)	< 0.001	19(90.5)	< 0.001	0	> 0.999	13(61.9)	< 0.001	12(57.1)	< 0.001
2	21(100)	20(95.2)		9(42.9)		2(9.5)		21(100)		8(38.1)		9(42.9)
3	0	0		0		0		0		0		0
Streak artifact
1	12(57.1)	10(47.6)	0.564	10(47.6)	0.564	11(52.4)	> 0.999	10(47.6)	0.564	10(47.6)	0.564	10(47.6)	0.705
2	8(38.1)	11(52.4)		11(52.4)		10(47.6)		11(52.4)		11(52.4)		11(52.4)
3	1(4.8)	0		0		0		0		0		0

Data are presented as the number of patients (n = 21) with percentages in parentheses

HIR, hybrid iterative reconstruction; DLR, deep learning reconstruction

Fig. 1.

Images of a 68-year-old patient (BMI 25.65 kg/m²) clinically diagnosed with ILD of unknown origin. (a) HRCT image reconstructed with HIR at standard settings. (b - d) LDCT image reconstructed with DLR (lung kernel) at mild, standard, or strong settings, respectively. (e - g) LDCT images reconstructed with DLR (bone kernel) at mild, standard, or strong settings, respectively. GGO, bronchiectasis, and bronchiolectasis can be seen in the lower lobes of both lungs (scored 1 or 2). The overall image quality was not significantly different between (a) and (b - g), but bronchiectasis and bronchiolectasis are slightly blurred in (g) compared with (a - f)

Assessment of normal and abnormal lung features

Most of the normal lung features were scored as 1 or 2. The LDCT DLR was comparable to the HRCT HIR in evaluating all normal lung structures (Table 4).

Table 4.

Assessment of normal lung features in HRCT HIR and LDCT DLR

Scores	HRCT HIR	LDCT DLR (lung, mild)	p	LDCT DLR (lung, standard)	p	LDCT DLR (lung, strong)	p	LDCT DLR (bone, mild)	p	LDCT DLR (bone, standard)	p	LDCT DLR (bone, strong)	p
Fissure
1	7(33.3)	9(42.9)	0.157	9(42.9)	0.157	9(42.9)	0.157	9(42.9)	0.157	9(42.9)	0.157	9(42.9)	0.157
2	14(66.7)	12(57.1)		12(57.1)		12(57.1)		12(57.1)		12(57.1)		12(57.1)
3	0	0		0		0		0		0		0
4	0	0		0		0		0		0		0
5	0	0		0		0		0		0		0
Proximal bronchi and adjacent pulmonary vessels
1	12(57.1)	14(66.7)	0.317	15(71.4)	0.18	15(71.4)	0.18	15(71.4)	0.18	15(71.4)	0.18	15(71.4)	0.18
2	8(38.1)	6(28.6)		5(23.8)		5(23.8)		5(23.8)		5(23.8)		5(23.8)
3	1(4.8)	1(4.8)		1(4.8)		1(4.8)		1(4.8)		1(4.8)		1(4.8)
4	0	0		0		0		0		0		0
5	0	0		0		0		0		0		0
Peripheral bronchi and adjacent pulmonary vessels
1	8(38.1)	9(42.9)	0.102	9(42.9)	0.102	9(42.9)	0.102	9(42.9)	0.102	9(42.9)	0.102	9(42.9)	0.102
2	10(47.6)	12(57.1)		12(57.1)		12(57.1)		12(57.1)		12(57.1)		12(57.1)
3	3(14.3)	0		0		0		0		0		0
4	0	0		0		0		0		0		0
5	0	0		0		0		0		0		0
Subpleural vessels
1	9(42.9)	10(47.6)	0.083	10(47.6)	0.083	10(47.6)	0.083	10(47.6)	0.083	10(47.6)	0.083	10(47.6)	0.083
2	9(42.9)	9(42.9)		9(42.9)		9(42.9)		9(42.9)		9(42.9)		9(42.9)
3	2(9.5)	2(9.5)		2(9.5)		2(9.5)		2(9.5)		2(9.5)		2(9.5)
4	1(4.8)	0		0		0		0		0		0
5	0	0		0		0		0		0		0

Data are presented as the number of patients (n = 21) with percentages in parentheses

HIR, hybrid iterative reconstruction; DLR, deep learning reconstruction

In total, 126 lung lobes were assessed for abnormal lung structures, including GGOs (82, 65.1%), reticulation (39, 31.0%), bronchiectasis and/or bronchiolectasis (50, 39.7%), architectural distortion (28, 22.2%), and honeycombing (8, 6.3%). LDCT DLR (bone, strong) visualization of bronchiectasis and/or bronchiolectasis was significantly weaker than that of HRCT HIR (p = 0.046) (Fig. 1). The assessment of other abnormal lung structures assessment showed no significant differences between LDCT DLRs and HRCT HIRs (Table 5; Fig. 2).

Table 5.

Assessment of abnormal lung structures in HRCT HIR and LDCT DLR

Scores	HRCT HIR	LDCT DLR (lung, mild)	p	LDCT DLR (lung, standard)	p	LDCT DLR (lung, strong)	p	LDCT DLR (bone, mild)	p	LDCT DLR (bone, standard)	p	LDCT DLR (bone, strong)	p
GGO
0	44(34.9)	44(34.9)	0.317	44(34.9)	0.317	44(34.9)	0.317	44(34.9)	0.317	44(34.9)	0.317	44(34.9)	0.317
1	24 (19.0)	24 (19.0)		24 (19.0)		24 (19.0)		24 (19.0)		24 (19.0)		24 (19.0)
2	46(36.5)	48(38.1)		48(38.1)		48(38.1)		48(38.1)		48(38.1)		48(38.1)
3	12(9.5)	10(7.9)		10(7.9)		10(7.9)		10(7.9)		10(7.9)		10(7.9)
4	0	0		0		0		0		0		0
5	0	0		0		0		0		0		0
Reticulation
0	87(69.0)	87(69.0)	> 0.999	87(69.0)	> 0.999	87(69.0)	> 0.999	87(69.0)	> 0.999	87(69.0)	> 0.999	87(69.0)	0.317
1	0	0		0		0		0		0		0
2	17(13.5)	17(13.5)		17(13.5)		17(13.5)		17(13.5)		17(13.5)		18(14.3)
3	21(16.7)	21(16.7)		21(16.7)		21(16.7)		21(16.7)		21(16.7)		20(15.9)
4	1(0.8)	1(0.8)		1(0.8)		1(0.8)		1(0.8)		1(0.8)		1(0.8)
5	0	0		0		0		0		0		0
Bronchiectasis and bronchiolectasis
0	76(60.3)	76(60.3)	0.317	76(60.3)	0.317	76(60.3)	0.317	76(60.3)	0.317	76(60.3)	0.317	76(60.3)	0.046
1	5(4.0)	4(3.2)		4(3.2)		4(3.2)		4(3.2)		4(3.2)		2(1.6)
2	25(19.8)	26(20.6)		26(20.6)		26(20.6)		26(20.6)		26(20.6)		27(21.4)
3	18(14.3)	18(14.3)		18(14.3)		18(14.3)		18(14.3)		18(14.3)		19(15.1)
4	2(1.6)	2(1.6)		2(1.6)		2(1.6)		2(1.6)		2(1.6)		2(1.6)
5	0	0		0		0		0		0		0
Architectural distortion
0	98(77.8)	98(77.8)	0.317	98(77.8)	0.317	98(77.8)	0.317	98(77.8)	0.317	98(77.8)	0.317	98(77.8)	0.317
1	8(6.3)	9(7.1)		9(7.1)		9(7.1)		9(7.1)		9(7.1)		9(7.1)
2	15(11.9)	14(11.1)		14(11.1)		14(11.1)		14(11.1)		14(11.1)		14(11.1)
3	5(4.0)	5(4.0)		5(4.0)		5(4.0)		5(4.0)		5(4.0)		5(4.0)
4	0	0		0		0		0		0		0
5	0	0		0		0		0		0		0
Honeycombing
0	118 (93.7)	118 (93.7)	> 0.999	118 (93.7)	> 0.999	118 (93.7)	> 0.999	118 (93.7)	> 0.999	118 (93.7)	> 0.999	118 (93.7)	> 0.999
1	0	0		0		0		0		0		0
2	4(3.2)	4(3.2)		4(3.2)		4(3.2)		4(3.2)		4(3.2)		4(3.2)
3	4(3.2)	4(3.2)		4(3.2)		4(3.2)		4(3.2)		4(3.2)		4(3.2)
4	0	0		0		0		0		0		0
5	0	0		0		0		0		0		0

Data are presented as the number of patients (n = 21) with percentages in parentheses

HIR, hybrid iterative reconstruction; DLR, deep learning reconstruction

Fig. 2.

Images of a 62-year-old (BMI 23.42 kg/m2) diagnosed with CTD-ILD. (a) HRCT image reconstructed with HIR using standard settings. (b - d) LDCT image reconstructed with DLR (lung kernel) at mild, standard, or strong settings, respectively. (e - g) LDCT image reconstructed with DLR (bone kernel) at mild, standard, or strong settings, respectively. The visualization of GGO in image (b - g) was not significantly different from image (a) (black arrow)

Discussion

This study revealed that the DLR method could maintain the quality and reduce the noise of LDCT images compared with HRCT HIR in early ILD patients.

To eliminate the influence of gravitation-dependent density, prone position scanning was designed for better recognition of ILA in this study. The survey performed by European Society of Thoracic Imaging had reported that 59% of responders indicated that additional prone position scanning would be performed on demand in clinical practice [9]. The high radiation exposure of prone position scanning may be the main cause that restricts its clinical use. In this study, the SSDE of prone HRCT was 7.57 ± 1.33 mGy, which was much greater than the 5.24 ± 0.87 mGy of supine RDCT in ILD reported by Xu et al. [23] and similar to the 7.29 ± 1.45 mGy in supine HRCT reported in our previous study [26]. After combination with DLR, the radiation dose was reduced 63.7% in prone LDCT scans, which is much lower than that in chest RDCT.

DLR could reduce the image noise and keep image quality at low dose [11, 27]while HIR trades off image noise and spatial resolution at different radiation doses [28, 29] Several studies performed on coronary CTA, CTPA and abdominal ultrahigh-resolution CT revealed a concurrently significant reduction in the radiation dose and maintenance of image quality with DLR [11, 12, 30–32]. This study applied DLR to prone chest LDCT scans, and the results confirmed that DLR image quality was maintained at a lower dose than that of HRCT HIR.

To date, studies comparing images reconstructed with DLR and HIR in chest CT have been limited. Singh et al. compared standard-dose CT HIR and LDCT DLR images from 22 patients and found no significant differences in image quality or lung nodule detection between these image series [32]. Lenfant et al. applied DLR and HIR (AIDR3D) reconstruction techniques to 45 and 48 CTPA images, respectively, and reported significantly greater image quality scores with DLR than with HIR. Nevertheless, the PE diagnostic confidence level showed no significant difference [12]. Similarly, our study of prone chest LDCT in early ILD patients revealed no significant difference in the overall image quality of the whole lung or lobar level between LDCT DLR and HRCT HIR.

Six DLR reconstruction algorithms (lung/bone, mild/standard/strong) were applied in this study. The objective assessment indicated that DLR significantly reduced image noise of LDCT, which was generally consistent with the subjective evaluation of image noise. Interestingly, the LDCT DLR (lung, mild) had slightly higher image noise than HRCT HIR, though the difference was not significant. No significant difference was found in the evaluation of normal or abnormal lung features between the two series of images. Nevertheless, the LDCT DLR (bone, strong) showed significantly less image noise than HRCT HIR, but the former was weaker for identifying in bronchiectasis and/or bronchiolectasis. According to previous studies, the bone kernel was recognized as a sharper, higher frequency kernel, as it was trained with high- and low-quality bone image pairs to produce an image quality similar to that of high-dose MBIR images. Several phantom studies have applied the bone kernel to small airway assessment and parenchymal density measurements and have shown improved spatial resolution and measurement accuracy [33, 34]. Additionally, small group tests in diffuse lung disease patients revealed that sharper images are helpful for clinical evaluation [35]. A phantom study by Greffier J. et al. demonstrated that image smoothness and feature detectability improved as the DLR levels increased [36]. Thus, the lung and bone kernels at the strong level were expected to produce the best image quality. However, our results were inconsistent with this expectation. Another DLR phantom study reported a difference in the behavior of the DLR strong level when comparing “body” and “body sharp” kernels. The researchers found that the mild and standard levels of the “body sharp” kernel better detected two different simulated lesions. In contrast, the strong level led to opposite changes in spatial resolution between high- and low-contrast lesions. The different spatial resolution results highlighted the potential uncertainty of “body sharp” kernels, and the authors conjectured that the training database of MBIR images might transfer nonlinear spatial resolution and noise characteristics to the DLR algorithm [14]. This finding might explain the opposite effects of “lung” and “bone” kernels in our study. Moreover, future studies with larger sample sizes and comparisons of multiple lesions should help clarify these findings.

This study has several limitations. First, as early ILD was usually occasional findings during the clinical practice, this study only enrolled a small number of patients, and more patients are needed to verify the results. Second, the scanning protocol was derived from a small-sample pretest in the supine position and was designed for equal image quality for diagnosis and a relatively lower radiation dose. Therefore, the maximal radiation dose reduction may not have been achieved, and further exploration of prone position scanning protocols is necessary, however, it should also be noted that the objective image noise of this protocol was relatively high, which might be related with the high BMI of enrolled patients, and researchers would better consider the image noise and radiation dose comprehensively. Third, the lack of unified clinical standards for prone position chest scans have led to selection bias, and a larger sample size and multicenter study may help reduce bias. Fourth, the image reconstruction methods applied in this study were specific algorithms of DLR and HIR, and the findings should be interpreted with cautiousness, and could not be directly generalized to other DLR technologies. The last but not least, the definition of ILA was subjective at present stage, and the relationship of ILA and clinical outcomes was still under study. Further studies could focus more on the features highly related to disease progression and clinical management.

Conclusion

DLR maintained image quality and reduced image noise in prone chest CT while reducing the radiation dose by 63.7%. LDCT reconstructed with DLR is potentially applicable for assessing early ILD in the prone position.

Abbreviations

ILD

Interstitial lung disease

DLR

Deep learning reconstruction

HIR

Hybrid iterative reconstruction

HRCT

High-resolution CT

LDCT

Low-dose CT

GGO

Ground glass opacity

DCCN

Deep convolutional neural network

PFTs

Pulmonary function tests

ILA

Interstitial lung abnormality

AIDR3D

Adaptive Iterative Dose Reduction 3-Dimensional

AiCE

Advanced Intelligence Clear-IQ Engine

VOIs

Spherical volume regions of interest

DLP

Dose-length product

CTDI_vol

Volume CT dose index

LAT

Lateral diameter

AP

Anteroposterior diameter

ED

Effective radiation dose

SSDE

Size-specific dose estimates

SD

Standard deviation

CTD

Connective tissue disease

Author contributions

RZ had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis, including and especially any adverse effects. RZ, YX, ZM contributed substantially to the study design, data acquisition, analysis and interpretation, drafting and revision of the manuscript. YW, JW, ZW, RX, YM, SP, JW, LS and PF contributed substantially to the data acquisition, analysis and interpretation, and revision of the manuscript. XS and WS both contributed to study concept and design, study supervision, acquisition of data, analysis and interpretation of data, revision of the manuscript, final approval of the version to be submitted.

Funding

This work was supported by National High Level Hospital Clinical Research Funding [grant numbers: 2022-PUMCH-A-034, 2022-PUMCH-B-069].

Data availability

Data is provided within the manuscript .

Declarations

Ethics approval and consent to participate

The Institutional Review Board of Peking Union Medical College Hospital approved this prospective study, and written informed consent was obtained from each participant.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.