Deep learning image reconstruction algorithm for carotid dual-energy computed tomography angiography: evaluation of image quality and diagnostic performance

This study was approved by the institutional review board at our institution. Written informed consent was obtained from all patients, and retrospective analyses were performed using a prospective cohort.

Participants

Forty-four consecutive patients who underwent carotid DECTA (Revolution CT; GE Healthcare that was able to reconstruct both ASIR-V and DLIR images) between November 2021 and December 2021 were included in the study. Among them 3 patients had metal implants after neck surgery; 5 patients had carotid stents; 5 patients had extensive thrombosis were excluded as accurate intravascular evaluations were not possible due to metal artifacts and thrombosis; and raw data of 3 patients were lost. Finally, 28 patients were included in this study.

Dual-energy CT technique

All examinations were performed by a fast kilovoltage-switching DECT scanner (Revolution CT; GE Healthcare) with carotid DECTA imaging parameters: 80/140 kV peak tube voltage, 12 HU noise index at 5-mm section collimation, variable tube current (GSI Assist; GE Healthcare); detector configuration, 128 detectors with 0.625 mm section thickness; 80 mm beam collimation, 0.5 s rotation time, 0.984:1 pitch, 36 cm display field-of-view. For contrast enhancement, nonionic-iodinated contrast agent (370 mgI/mL, Omnipaque 350, GE Healthcare, Shanghai, China) was injected intravenously into the right cubital at a rate of 3.5 mL/s using an automatic injector with a bolus of 40 mL and followed by 30-mL saline flush at the same injection rate. CTA was triggered by a bolus-tracking program (trigger point: the ascending aorta, trigger threshold: 120 HU) with a 5-s delay in image acquisition. CT images were obtained from the cranial crest to the aortic arch in the craniocaudal direction.

Imaging reconstruction

The raw data were reconstructed at 0.625 mm section thickness using 80% ASIR-V (IR 80% + FBP 20%) and DLIR algorithm at three selectable strength levels (low, DLIR-L; medium, DLIR-M; high, DLIR-H) graded by noise reduction capability. All reconstructions used standard kernel. All reconstructed images were processed into VMI at 50 keV using GSI Viewer software (Advantage Workstation, version 4.7, GE Healthcare).

Quantitative image analysis

Image analysis and measurement were performed on GSI Viewer software by one radiologist who was blind to images reconstruction algorithms. ROIs were manually placed on the axial images at levels of aortic arch (AA), common carotid artery (CCA), internal carotid artery (ICA), and vertebral artery (VA) at the dominating side to measure the mean attenuation (HU) and noise (SD) values. ROIs should cover the central parts of the arteries as much as possible while avoiding arterial wall, plaques, and severe artifacts. To calculate the CNR of the target vessel, HU values of sternocleidomastoid were also measured at the level of hyoid. Values of SNR and CNR were calculated using the equation: SNR = target HU / target SD and CNR = (target HU − muscle HU)/target SD.

Qualitative image analysis

In order to standardize qualitative analysis, two experienced radiologists were trained for image quality evaluation prior to qualitative image analysis. The VMIs dataset reconstructed by 80% ASIR-V, DLIR-L, DLIR-M, and DLIR-H were hanged in randomized order at GSI Viewer software to each radiologist without any annotations. All images were presented with preset window set: level, 100 HU and width, 800 HU, radiologists were allowed to adjust the window set during evaluations.

Two radiologists independently reviewed overall image quality for noise and texture of each image using a 5-point Likert scale: 5 = excellent for the best image quality, 4 = favorable (no influence on image interpretation), 3 = acceptable for diagnosis (possible influence); 2 = suboptimal (mild influence), and 1 = poor (impairing diagnosis). Two radiologists further rated the arterial depiction of head and neck artery using a 5-point Likert scale (Table 1) as published previously [23], according to vascular edge and subjective contrast to noise: 5 = very sharp edge with high contrast; 4 = sharp edge with satisfied contrast; 3 = minimal blurring edge with suboptimal contrast; 2 = blurring edge with markedly suboptimal subjective contrast; 1 = unacceptable blurring and contrast. In addition to CCA, ICA, VA and basilar artery (BA), we detailed the assessment of intracranial arteries, including anterior cerebral artery (ACA), middle cerebral artery (MCA), and posterior cerebral artery (PCA), at different segments according to Netter's cerebrovascular classification [24].

Table 1 Subjective image quality analysisDiagnostic evaluation

Moreover, two experienced radiologists who were blinded to the patient clinical data and images reconstruction algorithms were asked to evaluate the diagnostic performance to carotid plaques of the ASIR-V and DLIR images in consensus. Stenotic lesion was graded as the following criteria: 0 = no stenosis; 1 = mild (0–49%); 2 = moderate (50–69%); 3 = severe (70–99%); 4 = obstruction. According to composition variation, carotid plaques were divided into: 1, non-calcified plaques; 2, mixed plaque; 3, calcified plaques. There was an interval of 2 weeks between the evaluation of ASIR-V and DLIR images.

Statistical analysis

All statistical analyses were performed on software SPSS version 26.0 (IBM). Image quantitative parameters including HU, SD, SNR and CNR were analyzed using repeated measures ANOVA with the Bonferroni post hoc test between the ASIR-V, DLIR-L, DLIR-M, and DLIR-H groups. For the qualitative analysis including, overall image quality, and subjective ratings of different arterial segments, the Friedman test was conducted to compare these subjective indicators among four groups. The paired Wilcoxon signed-rank test was performed for post hoc subgroup comparisons when a significant difference was found between the four groups. The McNemar test was used to compare the diagnostic performance of DLIR and ASIR-V images. Inter-observe agreement and the agreement of the diagnostic results between DLIR and ASIR-V images were tested using Cohen’s kappa test, using the following criteria: poor (κ < 0.4); moderate (κ = 0.41–0.60); good (κ = 0.61–0.80); excellent (κ = 0.81–1.00). A p value < 0.05 was considered to indicate statistical significance.

留言 (0)

沒有登入
gif