Study design

This retrospective single-center study was approved by the Institutional Review Board of our institution, which waived the requirement for written informed consent due to the retrospective nature of the study.

We collected image data of patients who underwent contrast-enhanced CT venography of lower limbs. Consecutive patients with DVT lesions between September 2023 and September 2024 (n = 32) and those without DVT lesions between September 2023 and February 2024 (n = 20) were included (Fig. 1). One patient without DVT was excluded from the study due to reconstruction error.

Fig. 1

Patient inclusion flowchart. DVT, deep vein thrombosis

CT protocol

CT scans were performed using a multi-detector row CT (Aquilion One; Canon Medical Systems, Otawara, Japan). CT scanning parameters were as follows: scan mode: helical; tube voltage: 100 kVp; tube current: automatic tube current modulation with a standard deviation set at 13.0; helical pitch of 0.813:1. Iodinated contrast media at 600 mgI/kg was intravenously injected using a power injector. The injection rate was 20 mgI/kg/second (injection time: 30 s) and it was consistent across all patients. The CT venography was obtained 210 s after starting the injection. Images were reconstructed using the following algorithms from the raw data: DLR (Advanced intelligent Clear-IQ Engine with body sharp standard, Canon Medical Systems) and Hybrid IR (AIDR 3D enhanced standard with the kernel of FC03, Canon Medical Systems) and FBP with the kernel of FC03. The following image reconstruction parameters were consistent across all image sets: field of view: 350 mm (adjusted to the patient’s body size) and slice thickness/interval: 3/3 mm.

The Advanced Intelligent Clear-IQ Engine utilizes a vendor-trained convolutional neural network (CNN) architecture, specifically designed for noise reduction. The network was trained using a proprietary dataset comprising several clinical high-quality 120 kVp scans of the body in 10 patients. The validation dataset contained three independent image categories: patient images acquired at different dose levels, hepatic and pelvic phantom images obtained at varying dose levels, and images of metal objects and truncation artifacts [18].

Reference standard

Determination for correct DVT lesions was made through a consensus between one board-certified radiologist with 15 years of imaging experience and one radiology resident with 3 years of imaging experience.

Quantitative image quality analyses

Quantitative image quality analyses were conducted on the contrast-enhanced CT images by a radiology resident with 3 years of imaging experience using ImageJ (https://imagej.net/ij/). Circular or ovoid regions of interest (ROIs) measuring 3–5 mm in diameter were placed on the left femoral vein and subcutaneous fat on the same slice near the level of inguinal ligament (Fig. 2). The apparent lesion was avoided when placing regions of interest on these normal structures. The mean and standard deviation of the CT attenuation for each ROI were calculated and documented. The standard deviation of the CT attenuation for normal structures was considered quantitative image noise. The following metric were then computed:

Fig. 2

Circular or ovoid ROIs were delineated on the left femoral vein (black circle) and subcutaneous fat (white circle) on the same axial slice. ROI, region of interest

CNR = (MAVEIN − MAFAT)/√((SDVEIN² + SDFAT²)/2).

CNR, MAFAT, MAVEIN, SDFAT, and SDVEIN refer to the contrast-to-noise ratio, mean attenuation for subcutaneous fat and femoral vein, and standard deviation of the attenuation for subcutaneous fat and femoral vein, respectively. For the cases with DVT lesion, additional CNR between the thrombus and adjacent contrast-opacified vein lumen was calculated. The size of the ROI was selected so that it would not be larger than the size of the thrombus or vein lumen. The location and size of the ROIs were kept consistent between DLR, Hybrid IR, and FBP.

Lesion detection and qualitative image analyses

A board-certified radiologist with 15 years of imaging experience randomized all image sets. An additional radiologist (Readers 1, with 8 years of imaging experience) and two radiology residents (Reader 2 and 3, with 4 and 2 years of imaging experience, respectively) independently analyzed the contrast-enhanced CT images using ImageJ. They were blinded to the patient details and reconstruction methods.

The readers were asked to determine whether each patient had any DVT lesions and recorded confidence scores on a four-point scale (4, definitely present; 3, probably present; 2, uncertain for the presence or absence; and 1, no lesion). The test was conducted in three sessions, ensuring that DLR, Hybrid IR, and FBP images for the same patient did not appear in the same session. A 2-week interval was maintained between sessions to prevent recall bias.

The three same readers assessed the contrast-enhanced CT images at a minimum of 2 weeks after the lesion detection test. The images were assessed using the following criteria:

DVT lesion depiction (all lesions were annotated and assessed) (4, clear depiction; 3, slightly blurred; 2, moderately blurred; and 1, unrecognizable).

Structure depiction (veins and muscles) (4, clear depiction; 3, slightly blurred; 2, moderately blurred; and 1, unrecognizable).

Subjective image noise (4, less noise; 3, standard noise; 2, more than standard noise; and 1, severe noise).

Presence of artifacts (3, almost no artifacts; 2, mild artifacts; 1, severe artifacts affecting imaging diagnosis).

Overall image quality (5, excellent; 4, better than standard; 3, standard; 2, worse than standard; and 1, poor).

Statistical analysis

Statistical analyses were conducted using R (version 4.3.1; https://www.r-project.org/). The t-test and Fisher’s exact test were employed to compare demographic characteristics between the DVT lesion and non-DVT lesion groups. The paired t-test and Wilcoxon signed-rank test were utilized to compare the results for continuous variables and ordinal scales, respectively, between DLR and Hybrid IR, as well as between DLR and FBP. Diagnostic performance for lesion detection was calculated using the area under the receiver operating characteristic curve (AUC) and was subsequently compared with the DeLong test. To assess the sensitivity of the detection test, diagnostic confidence scores of ≥ 2 were considered indicative of lesion presence. These scores were evaluated statistically using McNemar’s test. Subgroup analyses of AUC and sensitivity were also performed based on lesion length (> 5 cm vs. <5 cm), lesion diameter (> 7 mm vs. <7 mm), location (femoral vs. peripheral), and occlusion grade (complete vs. partial) [19]. The interobserver agreement for lesion detection between the three readers was evaluated with the Fleiss’ kappa analysis. Kappa values of 0.00–0.20, 0.21–0.40, 0.41–0.60, 0.61–0.80, and 0.81–1.00 indicated slight, fair, moderate, substantial, and excellent agreement, respectively [20]. P-values < 0.025 were used to denote statistical significance after Bonferroni correction.