Patients. The institutional review board (IRB) of Samsung Medical Center approved this retrospective study (IRB, file number 2021-06-092), and the requirement of informed consent was waived for the use of patient medical data. All methods were performed in accordance with the Declaration of Helsinki. In total, 369 consecutive patients were included between August 2021 and September 2021. A CT scan was performed because of clinical or radiological suspicion of ILD on chest radiography. All patients underwent routine CT evaluation for ILD, which included helical LDCT and standard-dose non-helical high-resolution CT (HRCT). A total of 170 patients without imaging findings suggestive of ILD (for example, non-dependent ground-glass opacity or reticular abnormalities, non-emphysematous cysts, honeycombing, and traction bronchiectasis)29 were excluded. Six patients were excluded due to suboptimal image quality. Finally, 193 patients were included in this study.
CT acquisition and image reconstruction. All CT images were obtained using a multidetector CT scanner (Revolution Frontier, GE Healthcare) under the LDCT protocol without the use of contrast. The protocols consisted of a fixed tube current of 20 mAs per slice (40 mA with a half-second rotation and 0.984 pitch). Slice thickness of 1.25 mm and high-spatial-frequency algorithm were applied. The chest CT protocol used the helical mode with the following parameter: 1.25 mm × 64 detector configuration. The other parameters were as follows: peak tube voltage of 120 kVp, 40-mm table feed per gantry rotation, pitch of 0.984:1, and z-axis tube current modulation. Supine inspiratory HRCT scans of all patients were obtained without intravenous contrast using the same CT scanner. The protocol consisted of sections reconstructed with a high-spatial-frequency algorithm at 1- or 2-cm intervals under automatic exposure control (142–275 mA with dose modulation) with a slice thickness of 1.25 mm, from apex to base.
Three different reconstructions of the LDCT images were obtained: conventional FBP, adaptive statistical iterative reconstruction-Veo at a level of 30% (ASiR-V), and DLM. All scan data were directly displayed on the picture archiving and communication systems (PACS) (Centricity 3.0, GE Healthcare) workstation monitors. Images were viewed on monitors in lung (width, 1500 HU; level, -700 HU) window settings. To assess radiation exposure in LDCT, we reviewed the CT dose index (CTDIvol) and dose-length product (DLP) recorded as digital imaging and communications in medicine data. The estimated effective dose was calculated as the DLP multiplied by a k-factor of 0.014 mSv·mGy− 1·cm− 130.
Deep learning reconstruction model. The deep learning model reconstruction (DLM; ClariCT.AI, ClariPi)13 was developed as a denoising solution using a U-Net-based convolutional neural network. Details are summarized in Supplementary material (Supplementary Fig. S1 and S2).
Image quality analysis. The performance of the image reconstruction methods was evaluated both quantitatively and qualitatively for each case, reconstructed using three different methods (FBP, ASiR-V, and DLM). Quantitative analysis was performed using signal, noise, signal-to-noise ratio (SNR), and blind/referenceless image spatial quality evaluator (BRISQUE) score. For qualitative analysis, a thoracic radiologist visually scored the images.
Signal, noise, and signal-to noise ratio. Standardized 20-mm-diameter circular regions of interest were used to record signal and noise, which represented the mean pixel intensity value and standard deviation of pixels for the lung parenchyma, and background air for LDCT scans in FBP, ASiR-V, and DLM image sets25. Lung measurements were obtained from the lower lobes towards the periphery to avoid parenchymal lesions. Background air was obtained from the air external and anterior to the patient at the sternomanubrial junction31. The signal-to-noise ratio (SNR) was calculated for all three image sets. SNR was calculated as follows:
𝑆𝑁𝑅𝑏𝑎𝑐𝑘𝑔𝑟𝑜𝑢𝑛𝑑 𝑎𝑖𝑟= |(𝑠𝑖𝑔𝑛𝑎𝑙𝑏𝑎𝑐𝑘𝑔𝑟𝑜𝑢𝑛𝑑 𝑎𝑖𝑟)/(𝑛𝑜𝑖𝑠𝑒𝑏𝑎𝑐𝑘𝑔𝑟𝑜𝑢𝑛𝑑 𝑎𝑖𝑟)|, 𝑆𝑁𝑅lung p𝑎𝑟𝑒𝑛𝑐ℎ𝑦𝑚𝑎= |(𝑠𝑖𝑔𝑛𝑎𝑙lung p𝑎𝑟𝑒𝑛𝑐ℎ𝑦𝑚𝑎)/(𝑛𝑜𝑖𝑠𝑒lung p𝑎𝑟𝑒𝑛𝑐ℎ𝑦𝑚𝑎)|32.
BRISQUE. BRISQUE is a no-reference image quality assessment model that uses natural scene statistics in the spatial domain23. This model is composed of three steps: 1) extraction of natural scene statistics, 2) calculation of feature vectors, and 3) prediction of the image quality score. We utilized a pre-trained prediction model provided by Mittal et al.23 to predict the image quality score. The minimum and maximum image scores are 0 and 100, respectively, with a lower image score indicating a better image quality. The potential of BRISQUE as an indicator of medical image quality has been reported previously23,33,34.
Visual scoring. A radiologist with six years of experience in thoracic imaging performed a qualitative image analysis on a chest CT scan. The radiologist was blinded to the patients’ data and the image reconstruction techniques and examined the images in a random order using PACS. CT scans were graded on axial images with datasets displayed on standard windows, and windowing was allowed, as in routine reporting conditions. The reader randomly assessed the subjective image contrast, noise, and image quality using a five-point visual scoring system20 (Table 4).
Table 4
Visual scoring system used to evaluate image quality
Scale | Contrast | Noise | Overall image quality |
5 | Excellent | Minimal | Best |
4 | Above average | Below average | Slight inferior (no influence on diagnosis) |
3 | Acceptable | Average | Mildly inferior (possible influence on diagnosis) |
2 | Suboptimal | Above average | Moderately inferior (probable influence on diagnosis) |
1 | Poor | Unacceptable | Markedly inferior (impairing diagnosis) |
Evaluating diagnostic efficacy of ILD. Two thoracic radiologists (reader 1 had 6 years of experience; reader 2 had 15 years of experience) who were blinded to clinical data and image reconstruction techniques independently assessed CT images and determined the radiologic features of usual interstitial pneumonia (UIP), which is a hallmark of idiopathic pulmonary fibrosis (IPF). A classification of ‘UIP’, ‘Probable UIP’, ‘Indeterminate for UIP’, and ‘Alternative diagnosis’ was assigned for each case using the 2022 American thoracic society and Fleischner Society guidelines4,18,24. The UIP pattern was defined as subpleural, basal predominance of reticular abnormalities, honeycombing with, or without traction bronchiectasis; the absence of findings was suggestive of alternative diagnosis, including extensive ground-glass opacity, micronodules, discrete cysts, mosaic attenuation, or segmental/lobar consolidation24. The two readers formed a consensus diagnosis for each case reconstructed using three different methods (FBP, ASiR-V, and DLM) after independent assessment. After the final diagnosis was made using the three reconstruction methods, cases showing discrepant diagnoses were selected and compared with HRCT findings of the patient, which was considered a reference standard.
Statistical analysis. Image quality of the three reconstruction methods (FBP, ASiR-V, and DLM) were compared using one-way analysis of variance, and post-hoc pairwise comparisons were adjusted for multiple comparisons using the Bonferroni correction. Cohen’s kappa statistics were used to evaluate the agreement between the two readers and the diagnostic agreement on the three reconstruction methods. A kappa statistic of 0.81–1.00 indicates an excellent agreement; 0.61–0.80, substantial agreement; 0.41–0.60, moderate agreement; 0.21–0.40, fair agreement; and 0.00–0.20, poor agreement35. Statistical significance was set at p < 0.05. All statistical calculations were performed using SAS (version 9.4; SAS Institute, Cary, NC, USA) and R (version 3.3.1; Vienna, Austria; http://www.R-project.org/36) software.