High-Resolution Head CTA: A Prospective Patient Study Comparing Image Quality of Photon-Counting Detector CT and Energy-Integrating Detector CT

Patient Selection and CTA Technique for Both EID-CT and PCD-CT

Institutional review board approval was obtained for this Health Insurance Portability and Accountability Act–compliant prospective research study. Clinically indicated adult patients underwent head CTA with EID-CT (Force, Siemens Healthineers). Soon after the clinical EID-CT, at a time interval of approximately 30 to 60 minutes, with informed consent, a research head CTA was performed by using a PCD-CT (NAEOTOM Alpha). An identical injection of iodinated contrast was used for each examination: each of the 2 injections consisting of 100 mL Omnipaque 350 (GE Healthcare) were administered at 4 mL/s, followed by 35 mL saline at 4 mL/s. The radiation dose associated with the research PCD-CT scan was assessed by our board-certified medical physicists, and reviewed by our institute’s Radiation Safety Committee as well as our institutional review board. Similarly, the risks of a second contrast injection were also reviewed by our departmental and institutional committees, and finally approved by our institutional review board. The potential risk of the research PCD-CT scan, from both radiation and contrast agent perspectives, was deemed to be low. Verbal and written informed consent were obtained from all patients who participated in this study. During this process, all relevant information about the study, including potential risks, was communicated with the patients who were given time to make their decision on participation. A dual-energy scanning protocol (the clinical default) was used on the EID-CT (100/150Sn, 192 × 0.6 mm detector configuration, rotation time 0.5 second). A UHR scanning mode was used on the PCD-CT (120 kV, 120 × 0.2 mm). Allowing for the inherent differences between EID-CT and PCD-CT, the techniques were matched as much as possible, including with regard to dose. Specifically, the volume CT dose index (CTDIvol) was approximately matched.

Image Reconstruction and a Deep Convolutional Neural Network for Noise Reduction

To provide a fair comparison between the 2 CT systems, images from both platforms were reconstructed at 0.6-mm section thickness by using a closely matched medium-sharp SR kernel; Quantitative regular (Qr) 54 on EID (EID-SR) and Qr56 on PCD (PCD-SR), both with an iterative reconstruction strength setting of 3 and both by using a matrix size of 1024 × 1024. In addition, PCD-CT acquired with the UHR mode (120 × 0.2 mm) provides a much higher spatial resolution. To utilize the UHR capability on PCD-CT, we also reconstructed HR images with the sharpest quantitative kernel (Qr89) at the minimum achievable section thickness of 0.2 mm. This kernel and section thickness are not achievable on the EID-CT.

With this sharp reconstruction and thin section thickness on PCD-CT, images are very noisy. Therefore, a deep CNN previously developed and validated at our institution was applied to the UHR image series to reduce image noise. The CNN model in this study was specifically trained for UHR head CTA examinations on PCD-CT (PCD-UHR-CNN) (Fig 1). It used a modified 6-layer residual U-Net with 7 slices as input.^12,13 The training was performed by using pairs of input and target images that were created without and with the iterative reconstruction, respectively, both reconstructed with the Qr89 kernel and 0.2-mm section thickness. Eight cases were used in training and 2 cases used in validation. A large number of patches (256,000) with size of 128 × 128 × 3 was generated to alleviate the overfitting during training. In addition, both batch normalization and drop-out layers were added after convolution layers of the UHR-CNN model to ensure the model trained without overfitting. The training was performed on an NVIDIA Tesla M40 GPU with 12-GB memory.

• Download figure
• Open in new tab
• Download powerpoint

FIG 1.

Data preparation and training mechanism of a dedicated ultra-high resolution convolutional neural network (UHR-CNN) for denoising high resolution images from PCD-CT. The model used a modified 6-layer residual U-Net with 7 slices as input. The high-quality images were from iterative reconstruction (IR) images with a thick section thickness and the low-quality images were generated by adding noise patches from subtraction of filtered back-projection (FBP) and IR images followed by spatial decoupling.

Detailed scan and reconstruction parameters are shown in Table 1.

Image Quality Review Sessions and Evaluation Criteria

Two neuroradiologist readers (R1 and R2, one with 26 years and the other with 5 years of clinical practice experience after neuroradiology fellowship) independently evaluated image quality in a randomized and blinded fashion in 2 separate sessions: 1) PCD-SR versus EID-SR; and 2) PCD-UHR-CNN versus EID-SR. The purpose of the first session was to compare images from PCD-CT and EID-CT at matched spatial resolution. The purpose of the second session was to evaluate the UHR-CNN images from PCD-CT at a resolution not achievable on the EID-CT. The sessions were separated by at least 14 days. In both sessions, each reader evaluated the 2 series of images displayed side-by-side.

Overall image quality was assessed for each technique (5-point Likert score, ranging from 1 [nondiagnostic] to 5 [excellent image quality, better than routine]). The following image quality categories were also assessed (5-point Likert score, ranging from −2 [significantly inferior] to 2 [significantly superior]): noise, artifact, and visibility of 7 representative intracranial arterial segments on each patient’s right side (cavernous ICA, ophthalmic, M3 [opercular], A1, A3 [pericallosal], P1, and P3 [quadrigeminal]). Only these 7 segments and only the right side was chosen to mitigate against reader fatigue. The anterior and posterior communicating arteries were not chosen because they may not be well visualized in any given patient. Specific definitions of these segments included: M3 (opercular), between the insular and cortical segments, beginning at the top of the Sylvian fissure, coursing inferolaterally through it, and exiting at the surface of the cerebrum; A3 (pericallosal), arising from the A2 segment near the genu of the corpus callosum as the larger of the 2 major anterior cerebral artery branches, and overlying the corpus callosum and beneath the cingulate gyrus; and P3 (quadrigeminal), between the ambient and calcarine segments in the quadrigeminal cistern, extending posterior to the midbrain.

In addition, readers evaluated their diagnostic confidence of any specific clinically meaningful arterial findings (eg, stenosis, aneurysm, aneurysm versus infundibulum, etc), if present. All such findings were included, regardless of side (right, left, or midline). The readers were not asked to locate arterial findings as unknowns, nor were they asked to assess if one CT methodology could find more arterial findings than the other. Rather, any clinically meaningful arterial findings, if present, had previously been identified for the 2 readers by a third neuroradiologist (13 years of clinical practice experience after neuroradiology fellowship). This reviewer did not otherwise participate in the comparative reading sessions, and used the prior clinical reports from the EID-CT examinations and confirmatory EID-CT image review to identify the findings. Only the prior EID-CT reports and images were used to identify these findings, not any prior or subsequent CTA/MRA/DSA. A sliding scale was used for this diagnostic confidence assessment by R1 and R2: any number between 0 and 100, ranging from 0, finding extremely difficult to see; 25, finding somewhat difficult to see; 50, finding moderately well seen; 75, finding well seen; to 100, finding extremely well depicted. If a stenosis was preidentified by the third neuroradiologist, stenosis severity (5-point scale, 1: minimal 1%–24%; 2: mild 25%–49%; 3: moderate 50%–69%; 4: severe 70%–99%; 5: occluded) was also rated for each individual series. The 2 readers also were asked to provide subjective comments for any lesion, if relevant.

Iodine Contrast, Noise, and Contrast to Noise Ratio Measurements

Contrast enhancement of iodine was measured as the mean CT number of an ROI within the artery and subtracted from the background CT number measured in another ROI at a nearby uniform soft tissue area. Image noise was measured on the ROI of the soft tissue as the standard deviation of CT numbers for EID-SR, PCD-SR, and PCD-UHR-CNN. The ROI size used best fit to contain the given artery while avoiding the vessel wall due to potential blooming from calcification. The same ROI size was used across all modalities for each specific artery at 4 locations: right and left distal ICA and right and left midcervical ICA. Iodinated contrast to noise ratio (CNR) was calculated by averaging the values of these 4 locations for each patient and for all 3 imaging conditions: EID-SR, PCD-SR, and PCD-UHR-CNN.

Statistical Analysis

Statistical analysis was performed based on the scores obtained from both reader sessions. Means and standard deviations on patient characteristics, CT dose, and each reader’s scores for image quality category were calculated. For the image quality comparisons, a 2-tailed paired t test was used when appropriate, with a P value < .05 indicating a statistically significant difference. Image noise and CNR were also compared among the 3 image conditions with the Friedman test being used for all 3 image sets, with P value < .05 indicating a statistically significant difference. The average noise for PCD-SR and EID-SR was used for calculation of an absolute percent difference. Horizontal percentage stacked bar chart plots were generated to display reader evaluation for image quality, noise, artifacts, and visualization of small arterial segments (Fig 2). Bars are split into segments and the length of each bar is 100%. Grouped boxplots were generated for diagnostic confidence, noise, and CNR results (Figs 3 and 4). The bottom, top, and middle lines of each box indicate the 25th percentiles, 75th percentiles, and sample median of the data. Samples whose notches do not overlap are statistically significantly different at the 5% significance level (Matlab 2021, MathWorks). For the diagnostic confidence data comparing PCD-SR and PCD-HR-CNN with EID-SR, interrater intraclass correlation coefficients (ICC) were calculated. The ICC values were interpreted as follows: <0.5, poor; 0.5–0.75, moderate, 0.75–0.9, good; >0.9, excellent.

• Download figure
• Open in new tab
• Download powerpoint

FIG 2.

Reader evaluation, PCD-CT versus EID-CT. Horizontal percentage stacked bar chart plots demonstrate reader 1 and 2 (R1, R2) scores for PCD-SR versus EID-SR (A) and PCD-UHR-CNN versus EID-SR (B). Bars are split into segments and the length of each bar is 100%. Segments from left to right: significantly superior (+2), somewhat superior (+1), equal quality (0), somewhat inferior (−1), significantly inferior (−2).

• Download figure
• Open in new tab
• Download powerpoint

FIG 3.

Data plot of diagnostic confidence of arterial findings. Grouped boxplot figure highlighting PCD-SR versus EID-SR and PCD-UHR-CNN versus EID-SR for diagnostic confidence scores of each reader. The bottom, top, and middle lines of each box indicate the 25th percentiles, 75th percentiles, and sample median of the data. Samples whose notches do not overlap are statistically significantly different at the 5% significance level. (Matlab 2021).

• Download figure
• Open in new tab
• Download powerpoint

FIG 4.

Iodine contrast, noise, and CNR quantitative evaluation. Grouped boxplots for (A) iodine contrast, (B) noise, and (C) iodine CNR results measured in the bilateral distal ICA and midcervical ICA for EID-SR, PCD-SR, and PCD-UHR-CNN. The bottom, top, and middle lines of each box indicate the 25th percentiles, 75th percentiles, and sample median of the data. Samples whose notches do not overlap are significantly different at the 5% significance level. (Matlab 2021).