Application of deep learning reconstruction of ultra-low-dose abdominal CT in the diagnosis of renal calculi

Study population

This prospective study was approved by the Medical Ethics Committee of our institution (No. HS-2427). We obtained written informed consent for both abdominal LDCT and ULDCT from each enrolled patient. Patients with suspected renal calculi were recruited from November to December 2020. All patients underwent Canon CT scans. Exclusion criteria were age < 18 years and a history of abdominal or pelvic implantation, such as an arterial stent or artificial hip joint. No patients were excluded.

Image acquisition and reconstruction

The Aquilion ONE Genesis CT system (Canon Medical) was used to acquire images. The Rotation speed was 0.5 s/round, pitch was 0.813, and scan area ranged from the apex of the liver to the bifurcation of the bilateral common iliac arteries. The scan voltage used for both low and ultra-low doses was 100 kV. Tube current were adjusted automatically. The low-dose noise index was the standard setting (7.5), and the ultra-low-dose noise index was the low-dose setting (20). LDCT images were reconstructed with hybrid iterative reconstruction (HIR, Adaptive Iterative Dose Reduction 3-Dimensional, [AIDR3D]) (LD-HIR). ULDCT images were reconstructed with HIR (ULD-HIR) and DLR (ULD-DLR). Five-millimeter images were uploaded to the picture-archiving and communication system (PACS) for unified analysis.

Detection of calculi and abdominal lesions

A radiologist with 3 years’ working experience recorded the numbers and positions of renal calculi in images of all 3 groups and measured the diameter of each renal calculus twice to calculate a mean value. The radiologist was blinded to the groupings of images, and the images were presented in random order. LD-HIR image was used as the reference to calculate the renal-calculus detection rate from the ULD-HIR and ULD-DLR images. The same radiologist evaluated types and numbers of lesions in solid abdominal organs and measured their diameters twice to calculate the means.

Objective evaluation of image quality

Radiologist A performed quantitative analysis of cross-sectional images (section thickness, 5 mm). To measure CT value, noise (SD; standard deviation of the CT value) and signal-to-noise ratio (SNR; mean attenuation/SD), we delineated regions of interest (ROIs) in the liver, spleen, aorta, both kidneys, anterior abdominal-wall subcutaneous fat, and right psoas muscle on LD-HIR, ULD-HIR, and ULD-DLR images (Fig. 1). ROIs were kept consistent across all three groups. Those ROIs on the liver, spleen, and aorta were at the level of the hepatic hilum; those on both kidneys were at the level of the renal hilum; and the ROIs of the right psoas muscle and anterior abdominal-wall subcutaneous fat were at the level of the fourth lumbar vertebra. We placed 4 and 2 ROIs in the liver and kidneys, respectively, and calculated the means. One ROI was placed in the spleen, aorta, right psoas muscle, and subcutaneous fat each, and triplicate measurements were taken to calculate the means. ROI size for both kidneys was maintained at 0.4–0.5 cm2, while ROI size for all other organs was maintained at 0.8–1.0 cm2. Contrast-to-noise ratios (CNRs) for the liver, spleen, kidneys, and aorta were calculated by the following formula:

$$}_}}} = \left( }_}}} - }_}} } \right)/}$$

where CTorgan is the CT value of the organ of interest, CTpsoas muscle is the mean CT value of the right psoas muscle, and total image noise is the SD of subcutaneous fat in the anterior abdominal wall [27].

Fig. 1figure 1

ROIs were placed on the liver, spleen, aorta, kidneys, right psoas muscle, and abdominal-wall subcutaneous fat to measure CT values and noise of various abdominal organs and to evaluate the objective image quality. a shows ROIs located in the liver, spleen, and aorta. b shows ROIs located in the kidneys. c shows ROIs located in the right psoas muscle and anterior abdominal-wall subcutaneous fat

Subjective evaluation of image quality

Radiologists A and B, who had, respectively, 3 and 8 years of working experience and were blinded to the image groupings, performed 5-point Likert scoring of LD-HIR, ULD-HIR, and ULD-DLR images that were presented to them in random order. Scoring criteria were as follows: (1) extremely poor image quality, rendering diagnosis impossible; (2) poor image quality with serious noise; (3) medium image quality, sufficient contrast, and some noise; (4) good image quality, good contrast, and little noise; and (5) excellent image quality, good contrast, and no significant noise. The initial window width and position were set at 350 and 50 HU, respectively, and both parameters were modifiable.

Radiation dose

To evaluate radiation dose, we recorded the volume CT dose index (CTDIvol) and dose length product (DLP) on the scanner and calculated the effective radiation dose. The effective radiation dose corresponded to the value of the DLP multiplied by the abdominal conversion coefficient, which was 0.015 mSv/mGy.cm.

Statistical analysis

All analyses were performed using SPSS version 25.0 (IBM Corp., Armonk, NY, USA). The Kolmogorov–Smirnov test was used to determine whether the data were normally distributed. LD-HIR images were used as references to assess the image quality, renal-calculus measurements, radiation exposure, and lesion detection on ULDCT. We used the Kruskal–Wallis nonparametric test to analyze multigroup differences and the Mann–Whitney U nonparametric test for pairwise comparisons. p < 0.05 was considered statistically significant. We used Cohen’s weighted κ to calculate interobserver agreement, scored as almost perfect (0.81–1.00), substantial (0.61–0.80), moderate (0.41–0.60), fair (0.21–0.40), or poor (0.00–0.20).

留言 (0)

沒有登入
gif