Pushing the limits of spatial resolution in clinical PCD-CT using a dedicated high-resolution convolutional neural network (HR-CNN)

Abstract

Background:

Photon-counting-detector (PCD) CT systems offer ultra-high spatial resolution (up to 40 lp/cm), yet clinical practice is often constrained by pixel size, yielding resolutions below system capabilities. Even with a sharp reconstruction kernel, a large pixel size fails to fully utilize the kernel’s intrinsic spatial resolution. While reducing pixel size enhances resolution, it increases noise, compromising image quality.

Purpose:

To demonstrate the limit of spatial resolution in current and optimized implementations and maximize visual spatial resolution by addressing noise limitations.

Methods:

The relationship between spatial resolution, reconstruction kernel, and pixel size was investigated to identify strategies for utilizing the full spatial resolution potential of PCD-CT. To overcome the increased noise associated with high-resolution settings, a dedicated high-resolution deep convolutional neural network (HR-CNN) was developed to push the limit of visual spatial resolution in routine PCD-CT exams. The HR-CNN was trained using patient exams acquired with ultra-high-resolution (UHR) mode and reconstructed with a 150-mm field of view (FOV), matrix size of 1024×1024 (0.15-mm pixel size) and sharpest quantitative kernel (Qr89). The impact of FOV, kernel, and denoising on spatial resolution was studied using bar-pattern phantoms and a pilot clinical evaluation including 5 patients with interstitial lung diseases. Two thoracic radiologists evaluated 4 different FOV/reconstruction conditions: (1) FOV-410/Qr56-IR, (2) FOV-410/Qr89-IR, (3) FOV-150/Qr89-IR, (4) FOV-150/Qr89-HR-CNN in terms of overall image quality, noise, visual spatial resolution, and overall preference.

Results:

With a FOV of 410 mm, the Qr89 sharp kernel displayed bar-patterns up to 14 lp/cm, not much higher than the routine lung kernel Qr56. When the FOV was reduced to 150 mm, Qr89-IR allowed for the visualization of line pair patterns ranging from 18 to 20 lp/cm, with 20 lp/cm being moderately discernible. The application of Qr89-HR-CNN further improved this, enabling the display of line pair patterns as high as 20-22 lp/cm. In patient cases, both radiologists consistently ranked the FOV-150 images processed with HR-CNN as superior across metrics including overall image quality, noise reduction, visual spatial resolution, and overall preference. The HR-CNN reduced the noise in patients’ images by 93.0± 0.6% and 44.9± 5.3% in comparison with original FBP and IR (strength 4) images, respectively.

Conclusions:

Spatial resolution of PCD-CT is not fully utilized in routine practice due to the large FOV and high noise levels at sharp kernels. The proposed HR-CNN denoising method, along with small pixel size, may allow the high spatial resolution toward the system limit to be implemented in practice, which is beneficial in the diagnosis of many diseases, including interstitial lung disease.

1. Introduction

High spatial resolution CT imaging, enabled by recent scanner technologies including photon-counting-detector (PCD)-CT and energy-integrated-detector (EID)-CT with added grid/comb filters or reduced detector pixel sizes, plays an increasingly important role in many diagnostic tasks, including lung, musculoskeletal, inner ear, and cardiovascular exams.^1–4 Up to now, PCD-CT can achieve ultra-high in-plane spatial resolution up to 40 line-pairs per cm and 0.2 mm slice thickness using an ultra-high resolution (UHR) mode for scanning and a dedicated UHR kernel for image reconstruction.⁵ However, higher spatial resolution results in higher image noise. Without successful suppression of image noise, the benefit of high spatial resolution imaging may become quite limited.

The intrinsic in-plane spatial resolution in CT is determined by the x-ray focal spot, detector bin size, CT acquisition geometry, and reconstruction kernel. After reconstruction, the in-plane spatial resolution is also affected by the pixel size in the reconstructed image. In clinical practice, with a fixed CT acquisition geometry, in-plane spatial resolution is typically adjusted by selecting different reconstruction kernels. In addition, the selectable reconstruction field of view (FOV) and matrix size, which determine the pixel size in reconstructed images, also has a major impact on in-plane spatial resolution.

Using an ultra-sharp kernel without considering the combined effect of pixel size and noise may lead to suboptimal image quality. The spatial resolution of clinical images reconstructed from a sharp kernel is often below the intrinsic spatial resolution of the sharp reconstruction kernel if a standard matrix of 512×512 pixels and a large reconstruction FOV is used. In this situation, the spatial resolution of the image and the displayed structures is not limited by the reconstruction kernel, but by the pixel matrix of the image. To obtain a sufficiently small pixel size for a given reconstruction FOV, larger image matrices of 768×768, aresolution. However, reconstructing ultra-sharp, small-pixel images comes with the cost of high image noise. Noise increases with decreasing pixel size due to degrading photon statistics. In addition, high frequency noise will be present in the sharp kernel reconstructions that reconstruct the high frequency structures. Many radiologists preferred moderate sharp kernel for visualizing the appendicular skeleton,⁶ or even a soft kernel for hepatocellular carcinoma imaging ⁷ because sharp reconstructions incur increased image noise. Many studies investigated optimal noise-resolution trade-off for various clinical applications.^8–14

To optimize the trade-off between spatial resolution and noise, a direct solution is to reduce the noise associated with high spatial resolution reconstruction using noise reduction methods like iterative reconstruction (IR) and deep-learning-based image reconstruction (DLR). DLR algorithms typically learn the mapping function between low dose images and their high-dose counterpart during training.^15–20 It has been shown that DLR algorithms can effectively reduce image noise while maintaining the natural noise texture of FBP images, which could be ideal for noise reduction at high-resolution mode. Unfortunately, current PCD-CT systems on the market, which offer the highest spatial resolution, lack DLR methods to effectively reduce image noise, thus limiting the applications of high-resolution imaging.

There are two main purposes in this study. First, we will demonstrate the impact of reconstruction field of view, matrix size, and reconstruction kernel on spatial resolution using a state-of-the-art PCD-CT system and the limit of spatial resolution in current clinical implementations. Second, we will develop a dedicated high-resolution deep convolutional neural network (HR-CNN) to better utilize the high intrinsic spatial resolution of PCD-CT in clinical CT imaging.

2. Materials and Methods

2.1. Impact of pixel size on visual spatial resolution

System intrinsic spatial resolution is determined by CT focal spot, detector bin size, acquisition geometry, and reconstruction kernel. It is often measured and quantified via the pre-sampling modulation transfer function (MTF). Clinically, one factor that also significantly affects the visual impression of the spatial resolution (visual spatial resolution) is pixel size, defined by the field of view (FOV) divided by the reconstruction matrix size. For example, the high intrinsic spatial resolution (20 lp/cm) associated with a sharp kernel is not visible if the image is reconstructed with a 500 mm FOV and a 512×512 matrix, as the Nyquist frequency for this reconstruction is approximately 5.12 lp/cm, which is significantly lower than the kernel's intrinsic resolution. Here we distinguish intrinsic (objective) spatial resolution from visual spatial resolution (also termed perceived spatial resolution), which reflects the perceptual clarity of fine details as observed by a human reader in real images and is influenced by factors such as image noise, contrast, artifacts, and post-processing (e.g., denoising).²¹ A few combinations of reconstruction FOV and matrix size of PCD-CT are presented in Table 1. To demonstrate the effect of pixel size on visual impression of the spatial resolution, we varied the pixel size by using different reconstruction FOVs while fixing the matrix size at 1024.

Table 1.

Pixel sizes (with corresponding Nyquist limits in parentheses) for typical FOV/matrix size combinations

FOV (mm)╲Matrix size	512×512	768×768	1024×1024	2048×2048
150	0.30 mm (16.7 lp/cm)	0.20 mm (25.0 lp/cm)	0.15 mm (33.3 lp/cm)	0.07 mm (71.4 lp/cm)
280	0.55 mm (9.1 lp/cm)	0.36 mm (13.9 lp/cm)	0.27 mm (18.5 lp/cm)	0.14 mm (35.7 lp/cm)
410	0.80 mm (6.3 lp/cm)	0.53 mm (9.4 lp/cm)	0.40 mm (12.5 lp/cm)	0.20 mm (25.0 lp/cm)

A line pair phantom (high-contrast resolution module, advanced iqModules™, Gammex), featuring resolutions up to 32 lp/cm, was used to study the impact of reconstruction FOV and kernel on the visual impression of spatial resolution in PCD-CT. For reference, an in-house wire phantom in air (50 µm diameter, tungsten) was used to determine the system's limiting spatial resolution. This was achieved by measuring the pre-sampling MTF, which characterizes in-plane spatial resolution as a function of spatial frequency, using previously validated methods based on wire phantom images ²². The spatial frequency at 10% MTF, a standard value for quantifying image sharpness, was used to quantify the system’s limiting spatial resolution as a function of reconstruction kernel.

We performed visual spatial resolution assessment with regards to line pair phantom images for different kernel/reconstruction FOV conditions, and quantified the kernel-dependent spatial resolution using wire phantom images. Visual evaluations of line pair phantom images were performed by a board-certified medical physicist. To have a better visual assessment of each line pair, the window level/width was adjusted according to the mean CT value of a circular region of interest within the line pair region.

2.2. Dedicated high-resolution deep convolutional neural network (HR-CNN)

We developed an HR-CNN for high-resolution thoracic imaging on the PCD-CT scanner (NAEOTOM Alpha, Siemens Healthineers). The training/validation dataset contained 5,007 routine dose chest CT images reconstructed using both FBP and quantum iterative reconstruction (QIR) with a strength setting of 4, which was split into training and validation sets using 9:1 ratio. All images were reconstructed with the sharpest quantitative kernel (Qr89) and a pixel size of 0.15 mm (150-mm reconstruction FOV and 1024 matrix).

The developed training framework for the HR-CNN used image-based noise insertion to generate paired low- and high-noise CT images from the same patient dataset for supervised learning. The HR-CNN was trained using patient exams acquired with ultra-high-resolution mode. Implementation details are presented in the following steps, see Figure 1 for a schematic of the training mechanism.

(Step 1) Noise-only images were generated by taking the difference of thin-slice FBP and IR reconstructions with slice thickness/increment of 0.2/0.2mm at the same anatomic location. Thick-slice IR images (0.4/0.2mm) were reconstructed for step 3.

(Step 2) At the beginning of each training epoch, a random translation in the axial plane with the range of -15 to +15 pixels was applied to each noise-only image, which efficiently generated a sufficient amount of noise insertion examples and minimized the propagation of errors associated with the IR algorithm into the CNN.²⁰

(Step 3) Noise-only images from Step 1 were scaled by random factors (uniform distribution, 0.5–2.0) to vary noise intensity and improve model generalization. Thick-slice IR images (0.4 mm thickness, 0.2 mm increment) served as low-noise references. Both were cropped into 100,000 overlapping 3D patches (128 × 128 × 7 pixels) using a sliding-window approach with a stride of 64 pixels in-plane and 3 slices axially, ensuring efficient data augmentation through overlap while controlling computational demands.

(Step 4) Scaled noise-only patches were added pixel-wise to corresponding thick-slice IR patches to form high-noise multi-slice inputs. This approach simulated FBP noise characteristics while maintaining anatomical consistency, allowing the HR-CNN to learn cross-slice noise correlations essential for thoracic imaging.

(Step 5) Unaltered thick-slice IR patches were used as training targets, promoting noise reduction without detail loss.

(Step 6) Following training, the HR-CNN was directly applied to thin-slice FBP patient images (testing dataset) for the purpose of noise reduction.

Figure 1.

Training mechanism of the proposed high-resolution deep convolutional neural network (HR-CNN) denoising method. All training data was derived from patient image series reconstructed using FBP with thin-slice thickness (0.2mm) and QIR with both thin- and thick-slice thickness (0.2 and 0.4 mm), all with slice increment of 0.2 mm. Multiple-slice input was adopted to improve the performance of HR-CNN.

The proposed HR-CNN used a modified 6-layer residual U-Net as the architecture,²³ see Figure 2. The architecture consists of five max pooling and up-convolutional layers, 64 filters with the number of filters increased by 64 at each pooling layer, and exponential linear unit (ELU) activation. The network was trained for 200 epochs using Adam optimizer²⁴ with a descending learning rate from 0.001 to 0.00001, and a mini-batch of 16 image patches for each iteration. We employed pixel-wise mean-squared error between HR-CNN output and ground truth as the loss function during optimization. The HR-CNN training was performed on a NVIDIA Tesla M40 GPU with 12 GB memory.

Figure 2.

Schematic of the modified U-Net architecture used for this study. Conv.: convolutional layer, BN: batch normalization, ELU: exponential linear unit.

2.3. Clinical evaluation of HR-CNN at high resolution

The performance of the dedicated HR-CNN on visual spatial resolution of PCD-CT was studied with routine-dose chest CT images of five adult patients with interstitial lung diseases. The five patients were retrospectively selected from PCD-CT chest exams performed between March 22, 2023, and August 11, 2023, under an IRB-approved protocol. During this period, a total of 15 ultra-high-resolution (UHR) PCD-CT chest exams were performed on patients with known or suspected interstitial lung disease. The five cases were chosen as a sample based on (1) diagnostic confirmation of ILD, (2) image quality suitable for high-resolution analysis (absence of significant motion or artifacts), and (3) representation of typical ILD features (e.g., reticulation, septal lines, ground-glass opacities, traction bronchiectasis). Patient image data were collected and maintained following HIPAA-compliant and IRB-approved procedures. All enrolled participants were over 18 years of age and had consented to use their image data for research purposes. Patient image data were collected and maintained following HIPAA compliant and IRB approved procedures. Enrolled study participants were over 18 and consented to use of their image data for research purposes. Table 2 provides a summary of the scan/reconstruction parameters for the pilot clinical evaluation.

Table 2.

Scan and reconstruction parameters

Scan model	Siemens NAEOTOM Alpha (PCD-CT)
Scan mode	UHR
Detector collimation	120×0.2 mm
kV	120
Rotation time (s)	0.25
Helical pitch	1.0
CAREDose4D	on
CAREkeV IQ Level	70
CAREkeV optimized for	Non-contrast
Evaluation conditions: Recon FOV (mm)/Slice thickness (mm)/Kernel/Recon algorithm	410/0.8/Qr56/IR
	410/0.2/Qr89/IR
	150/0.2/Qr89/IR
	150/0.2/Qr89/HR-CNN

Clinical evaluation was based on visual assessment of patient images with different kernel/reconstruction FOV conditions. The impact of reconstruction FOV, kernel, and denoising on visual spatial resolution was investigated in a pilot study involving 5 patients with interstitial lung diseases. Three thoracic radiologists with 5, 6 and 14 years of thoracic radiology experiences evaluated 4 different FOV/reconstruction conditions: (1) FOV-410/Qr56-IR, (2) FOV-410/Qr89-IR, (3) FOV-150/Qr89-IR, (4) FOV-150/Qr89-HR-CNN in terms of overall image quality, visual spatial resolution, noise and preference for the diagnosis of interstitial lung disease (ILD), using Likert scales detailed in Table 3. Images processed by each of the four conditions were displayed simultaneously on a multiviewer in a single sitting in a darkened reading room. The radiologists, blinded to the conditions of the study, reviewed the axial images of all five cases. Noise was also measured as standard deviation of CT number at uniform regions of interest (aorta) to have a quantitative evaluation of noise reduction from the HR-CNN method.

Table 3.

Scoring system associated with Likert scales

Overall Image Quality (4-Point Scale)	Spatial Resolution (5-Point Scale)	Noise (5-Point Scale)	Preference for the Diagnosis of Interstitial Lung Disease (4-Point Scale)
1. non-diagnostic 2. diagnostic with low confidence 3. diagnostic with moderate confidence 4. diagnostic with high confidence	1. poor resolution not acceptable 2. acceptable spatial resolution but worse than routine 3.acceptable spatial resolution like routine 4. spatial resolution higher than routine diagnosis 5. superbly high spatial resolution	1. too noisy not acceptable 2. high noise but acceptable 3.acceptable noise like routine 4. low noise 5. very low noise	ranking order 1-4, with 1 the worst and 4 the best

3. Results

3.1. Impact of Pixel Size on Visual Spatial Resolution

As outlined in Table 1, pixel size in photon-counting detector CT (PCD-CT) varies with reconstruction FOV at a fixed matrix size of 1024, resulting in pixel sizes of 0.40 mm, 0.27 mm, and 0.15 mm for FOVs of 410 mm, 280 mm, and 150 mm, respectively. To evaluate the effect of pixel size on visual spatial resolution, Figure 3 presents a comparison of line pair images across these reconstruction FOVs using reconstruction kernels with an iterative reconstruction (IR) algorithm: Qr56, Qr64, Qr72, Qr80, Qr89, and Br98. The highest visible line pair pattern for kernels Qr56 and Qr64 remains at 10 lp/cm across all reconstruction FOVs, with no improvement in visual spatial resolution observed when using smaller reconstruction FOVs for these kernels. For Qr72 images, visual spatial resolution improves from 12–14 lp/cm to 14–16 lp/cm as the reconstruction FOV decreases from 410 mm to 280 mm; however, further reducing the reconstruction FOV to 150 mm yields no additional improvement. The advantage of smaller reconstruction FOVs becomes evident with sharper kernels. For kernel Qr80, visual spatial resolution improves from 14 lp/cm at a reconstruction FOV of 410 mm to 16–18 lp/cm at 280 mm. Further reduction to a 150 mm reconstruction FOV enhances resolution to 18–20 lp/cm, with the 20 lp/cm pattern being faintly discernible but noticeably clearer than at 280 mm. For the Qr89 kernel, the highest visible line pair pattern reaches 18–20 lp/cm at a reconstruction FOV of 150 mm. For the sharpest kernel, Br98, at a reconstruction FOV of 150 mm, the visual spatial resolution achieves 22–24 lp/cm, with higher line pair patterns such as 26 lp/cm faintly visible and even 28 lp/cm discernible with some structural details preserved. Conversely, when a large reconstruction FOV (e.g., 410 mm) is used, employing sharper kernels provides no significant improvement in visual spatial resolution. The benefits of sharper kernels are most pronounced with smaller reconstruction FOVs, such as 150 mm, where smaller pixel sizes enable visualization of finer line pair patterns.

Figure 3.

Line pair phantom (high-contrast resolution module, advanced iqModules™, Gammex) images reconstructed with a patient-size FOV, 410 mm, and smaller reconstruction FOVs of 280mm and 150 mm, each with Qr56(3), Qr64(4), Qr72(4), Qr80(4), Qr89(4) and Br98(4). Qr56 is our routine lung kernel; Qr89 is the sharpest quantitative kernel available on the PCD-CT; Br98 is the sharpest kernel with edge enhancement.

Figure 4 displays the presampling modulation transfer function (MTF) curves for reconstruction kernels using a filtered back projection (FBP) algorithm, illustrating system spatial resolution. The expected trend is evident: sharper reconstruction kernels yield higher system spatial resolution. The 10% MTF values, derived from these curves, were used to calculate the threshold pixel size and corresponding threshold reconstruction FOV at a fixed 1024 matrix size for each kernel. Exceeding the threshold pixel size for a given CT acquisition and reconstruction kernel compromises spatial resolution. Table 4 summarizes the kernel and reconstruction FOV selections based on MTF values and the highest visible line pair patterns for six reconstruction kernels: Qr56, Qr64, Qr72, Qr80, Qr89, and Br98.

Figure 4.

Presampling modulation transfer functions of six reconstruction kernels: Qr56, Qr64, Qr72, Qr80, Qr89 and Br98.

Table 4.

Kernel and FOV selection based on MTF and visible line pair pattern for reconstruction kernels

Reconstruction kernel	Br98	Qr89	Qr80	Qr72	Qr64	Qr56
Measured 10% MTF (lp/cm)	42	34	25	18	12	11
*Threshold Pixel size (mm)	0.12	0.15	0.20	0.28	0.43	0.46
*Threshold reconstruction FOV at 1024 matrix (mm)	120	150	200	280	430	460
Highest visible line pair pattern (lp/cm)	22-24	18-20	18-20	14-16	10	10

^*Threshold pixel size and reconstruction FOV, above which spatial resolution is compromised

Figure 5 presents example images from two patients, reconstructed with three reconstruction FOVs (410 mm, 280 mm, and 150 mm) and three reconstruction kernels—Qr56, Qr72, and Qr89—using an iterative reconstruction (IR) algorithm with slice thicknesses of 0.8 mm, 0.8 mm, and 0.2 mm, respectively. At a 410 mm reconstruction FOV (0.40 mm pixel size at 1024 matrix), no significant improvement in visual spatial resolution is observed when using the sharp kernel Qr89 compared to the routine lung kernel Qr56. At a 150 mm reconstruction FOV (0.15 mm pixel size), the Qr89 kernel enhances visualization of fine anatomical details, but this is accompanied by increased image noise. Larger pixel sizes, as with 410 mm and 280 mm reconstruction FOVs, often constrain visual spatial resolution below the capabilities of sharper kernels like Qr89. Conversely, employing an ultra-sharp kernel such as Qr89 without accounting for pixel size, noise, and the balance between visual and system spatial resolution may lead to suboptimal image quality.

Figure 5.

Example patient images (a, case 1, axial view; b, case 2, coronal view) with the 3 reconstruction FOVs of 410mm, 280mm and 150 mm, each with Qr56-IR, Qr72-IR, and Qr89-IR.

3.2. Impact of HR-CNN denoising

To mitigate the increased image noise associated with small pixel sizes and high-resolution reconstruction kernels, a dedicated high-resolution convolutional neural network (HR-CNN) model was applied to images reconstructed with a 150 mm reconstruction FOV and the sharpest quantitative kernel Qr89, using FBP algorithm. Figure 6 compares line pair images across six FOV/reconstruction/slice thickness conditions: (1) FOV 410 mm/Qr56-IR/0.2 mm, (2) FOV 410 mm/Qr89-IR/0.2 mm, (3) FOV 150 mm/Qr56-IR/0.2 mm, (4) FOV 150 mm/Qr89-IR/0.2 mm, (5) FOV 150 mm/Qr89-IR/0.4 mm, and (6) FOV 150 mm/Qr89-HR-CNN/0.2 mm. At a 410 mm FOV (0.40 mm pixel size), the Qr89 kernel resolves line pair patterns up to 14 lp/cm, comparable to the routine lung kernel Qr56. In contrast, combining a 150 mm reconstruction FOV with the Qr89 kernel achieves an in-plane spatial resolution of 18–20 lp/cm, markedly higher than configurations with a larger reconstruction FOV (410 mm) or smoother kernel (Qr56). The 18 and 20 lp/cm line pair groups are more clearly delineated with a 0.4 mm slice thickness (reconstruction FOV 150 mm/Qr89-IR/0.4 mm) compared to a 0.2 mm slice thickness, primarily due to reduced noise from the thicker slice, though this improvement sacrifices z-axis spatial resolution. By applying HR-CNN denoising (reconstruction FOV 150 mm/Qr89- HR-CNN/0.2 mm), in-plane spatial resolution is further enhanced to 20–22 lp/cm while preserving z-axis spatial resolution.

Figure 6.

Line pair phantom (high-contrast resolution module, advanced iqModules^™, Gammex) images reconstructed with a patient-size FOV, 410 mm, and a small reconstruction FOV, 150 mm, each with Qr56-IR and Qr89-IR. HR-CNN was also applied on the small reconstruction FOV sharp image to reduce noise. Arrows point to the highest visible bar-patterns.

Figure 7 displays example patient images reconstructed under four FOV/reconstruction/slice thickness conditions: (1) FOV 410 mm/Qr56-IR/0.8 mm, (2) FOV 410 mm/Qr89-IR/0.2 mm, (3) FOV 150 mm/Qr89-IR/0.2 mm, and (4) FOV 150 mm/Qr89-HR-CNN/0.2 mm. These results align with those observed in Figure 6. At a 410 mm FOV (0.40 mm pixel size at 1024 matrix), the sharp kernel Qr89 offers no significant improvement in visual spatial resolution compared to the routine lung kernel Qr56. At a 150 mm reconstruction FOV (0.15 mm pixel size), the dedicated high-resolution convolutional neural network (HR-CNN) model applied to Qr89 FBP images achieves superior image quality, combining improved visual spatial resolution with reduced noise compared to Qr89-IR images. Visual assessments by two radiologists indicate that septal lines (yellow arrowheads) and small vessels (orange arrowheads) are clearly delineated in HR-CNN images, whereas these structures appear as false ground glass opacities in other reconstructions. Additional patient images in Figure 8 further demonstrate the marked improvement in visual spatial resolution achieved with HR-CNN.

Figure 7.

Example patient images with the 4 FOV/reconstruction conditions: (1) FOV410/Qr56-IR, (2) FOV410/Qr89-IR, (3) FOV150/Qr89-IR, (4) FOV150/Qr89-HR-CNN.

Figure 8.

Two example cases show the visual spatial resolution improvement after applying a dedicated high-resolution deep convolutional neural network (HR-CNN) to the IR images with kernel Qr89.

Figure 9 displays another example patient images reconstructed under three FOV/reconstruction /slice thickness conditions: (1) FOV 410 mm/Qr56-IR/0.8 mm, (2) FOV 150 mm/Qr89-IR/0.2 mm, and (3) FOV 150 mm/Qr89-HR-CNN/0.2 mm. Visual assessments by two radiologists indicate that predominant abnormality appears as ground-glass opacity in the routine reconstruction (FOV410/Qr56-IR), while HR-CNN improves the ability to distinguish reticulation from ground-glass opacity, with clearer delineation of fine linear structures compared to noisy high-resolution IR (FOV 150 mm/Qr89-IR).

Figure 9.

Example patient images with the 3 FOV/reconstruction conditions: (1) FOV410/Qr56-IR, (2) FOV150/Qr89-IR, (3) FOV150/Qr89-HR-CNN. Predominant abnormality is ground glass opacity in FOV410/Qr56-IR, HR-CNN improves the ability to distinguish reticulation from ground glass opacity.

Five patient cases, each reconstructed under four FOV/reconstruction/slice thickness conditions—(1) FOV 410 mm/Qr56-IR/0.8 mm, (2) FOV 410 mm/Qr89- IR/0.2 mm, (3) FOV 150 mm/Qr89-IR/0.2 mm, and (4) FOV 150 mm/Qr89- HR-CNN/0.2 mm—were assessed by two experienced thoracic radiologists for visual spatial resolution, noise, overall image quality, and preference for diagnosing ILD. Subjective image quality scores, reported as means ± standard deviations (SDs) across all cases, are summarized in Table 5. Statistical analysis was performed using paired Student’s t-tests with a significance threshold of p < 0.05. No significant differences were observed between reconstruction FOV 410 mm images reconstructed with Qr56-IR/0.8 mm and Qr89-IR/0.2 mm for any of the four image quality criteria (p > 0.05). Images at reconstruction FOV 150 mm with Qr89-IR/0.2 mm scored significantly higher than both reconstruction FOV 410 mm conditions in visual spatial resolution, overall image quality, and preference for ILD diagnosis (p < 0.05), but exhibited significantly higher noise (p < 0.05). Across all cases, reconstruction FOV 150 mm images with Qr89-HR-CNN/0.2 mm were consistently ranked highest by both radiologists for visual spatial resolution, noise, overall image quality, and preference for ILD diagnosis (p < 0.05). Quantitative noise measurements, based on the standard deviation of CT numbers in the aorta for reconstruction FOV 150 mm images, revealed that HR-CNN reduced noise by 93.0 ± 0.6% compared to original FBP images and by 44.9 ± 5.3% compared to IR images.

Table 5.

Subjective image quality scores for different algorithms

	FOV-410/Qr56-IR/0.8mm	FOV-410/Qr89-IR/0.2mm	FOV-150/Qr89-IR/0.2mm	FOV-150/Qr89-HR-CNN/0.2mm
Visual spatial resolution*	2.90±0.32	2.70±0.48	3.90±0.32	4.50±0.53
Noise*	3.10±0.32	2.9±0.32	2.60±0.52	4.50±0.53
Overall image quality**	3.20±0.63	2.70±0.67	3.60±0.52	3.90±0.32
Preference on diagnosis of ILD**	2.10±0.57	1.20±0.63	2.70±0.48	4.00±0.00

^*1 is worst and 5 is best;

^**1 is worst and 4 is best

4. Discussion

This study investigated the combined effects of pixel size (determined by reconstruction FOV and matrix size) and reconstruction kernels on visual spatial resolution in photon-counting detector CT using phantom and clinical images. We demonstrated that, with a typical reconstruction FOV and matrix size in routine exams like thoracic CT, the high spatial resolution of the system is compromised even if a sharp kernel is used. To fully utilize the high-resolution capability of the system in clinical CT, a small pixel size (achieved by small reconstruction FOV and/or large matrix size) for reconstruction is required, which leads to drastically increased noise. To reduce noise at high spatial resolution, we proposed a dedicated HR-CNN denoising method in this study. Our results demonstrate that HR-CNN effectively extends visual spatial resolution toward the system limit of PCD-CT by mitigating the increased noise associated with small pixel sizes (e.g., 0.15 mm at 150 mm reconstruction FOV, 1024 matrix) and high-resolution quantitative kernels (e.g., Qr89).

Previous studies have shown that the high spatial resolution potential of PCD-CT is underutilized in routine practice.^8–14 Patzer et al concluded that spatial frequency limits alone are not sufficient for predicting the visualization capabilities in the appendicular skeleton with PCD-CT: fracture accessibility benefits from sharp non-ultrahigh-resolution (UHR) and moderate UHR kernels, while ultra-sharp reconstructions incur increased image noise with subsequently decreased discrimination of bone microarchitecture.⁶ Graafen et al investigated the impact of sharpness of kernel on image quality and concluded that soft reconstruction kernels yield the best overall quality for the evaluation of hepatocellular carcinoma in PCD-CT.⁷ For most of current PCD imaging applications, moderate UHR kernels are preferred over sharp UHR ones. For example, radiologists preferred PCD-CT images reconstructed with body sharp kernel Br76^8,9 or V71¹⁰ in musculoskeletal imaging, Q65¹¹ or Qr76¹² kernels for thoracic imaging, and Bv64/Bv72¹³ or B46¹⁴ kernels for cardiovascular imaging.

A previous deep learning denoising method for UHR-PCD applications, including skeletal surveys, lung screening, and head angiography, has been developed.²⁰ However, this prior work did not explore the relationship between spatial resolution and reconstruction settings (reconstruction kernels, FOV, and matrix size), limiting its ability to maximize PCD-CT’s spatial resolution potential. Although the Qr89 kernel was used in the previous study, reconstructions employed reconstruction FOVs of 280 mm or larger, which constrained the achievable resolution (16–18 lp/cm). In contrast, the current HR-CNN model specifically targets smaller pixel sizes (0.15 mm at 150 mm reconstruction FOV) to push the limits of visual spatial resolution of PCD-CT. The current HR-CNN includes several important improvements over the prior denoising approach²⁰:

1. Reference image generation: Unlike the previous model, which averaged five adjacent thin-slice IR images to create a thick reference target, HR-CNN uses directly reconstructed thick-slice IR images (0.4 mm thickness). Radiologist evaluations confirmed that these direct reconstructions yield superior visual spatial resolution compared to averaged thin-slice images (0.2 mm) at equivalent noise levels and slice thickness under the sharp Qr89 kernel.

2. With CAREDose4D enabled and consistent protocol parameters, noise in images reconstructed with the same kernel remains relatively comparable but varies ~10–30% across patients due to body habitus and anatomical location. The 0.5–2.0 scaling range was selected to accommodate this observed variability while addressing the training-inference noise mismatch introduced by thicker-slice IR targets, thereby promoting robust generalization.

3. Input dimensionality: HR-CNN employs multiple-slice inputs (128×128×7 pixels, 100,000 patches) instead of single-slice inputs (64×64 pixels, 50,000 patches), incorporating contextual information from adjacent slices to ensure z-axis structural continuity.

4. Spatial augmentation: The previous model employed randomized spatial decoupling with translations of 1–16 pixels and random sign inversion (multiplier of 1 or -1). In contrast, the HR-CNN model adopts a broader translation range of -15 to +15 pixels, including zero-pixel translations, while excluding inversion. Random sign inversion spatially decorrelates noise, resulting in a different noise texture than what is normally found in CT. The proposed HR-CNN approach enhances noise insertion diversity, mitigates error propagation in iterative reconstruction (IR) algorithms, and reduces false-positive structures in denoised outputs.

5. Dataset and resolution specificity: HR-CNN was trained on a larger dataset (5,007 routine-dose chest CT images) with the sharpest quantitative kernel Qr89 and ultra-high-resolution parameters (0.15 mm pixel size, 1024 matrix), tailored for thoracic PCD-CT applications to maximize visual spatial resolution.

These modifications enable HR-CNN to achieve noise reductions of 93.0 ±0.6% (vs. FBP) and 44.9 ± 5.3% (vs. IR) while enhancing visual in-plane resolution to 20–22 lp/cm, surpassing the performance of prior denoising approaches.

While HR-CNN substantially reduces image noise, makes high-resolution reconstruction clinically acceptable, and improves the visual spatial resolution (subjective perceptibility and diagnostic clarity of fine anatomical features) as evidenced by higher radiologist scores for perceived spatial resolution (Table 5), localized mild smoothing of high-contrast edges may occur in some structures compared to noisier IR reconstructions. But the net clinical benefit—reflected in superior overall image quality, noise, and diagnostic preference scores—outweighs this trade-off in the evaluated ILD cases.

Our training strategy utilizes routine-dose FBP-IR differences for noise insertion, which is practical as it requires no proprietary projection data access or external simulation tools and can be efficiently implemented in research or clinical environments. While this image-domain approach may introduce minor texture or resolution biases from the IR algorithm itself, several design choices in the framework effectively mitigate these risks and prevent error propagation or overfitting to reconstruction artifacts. In particular, the use of directly reconstructed thick-slice IR images (0.4 mm) as training targets provides inherently higher-quality, lower-noise references compared to thin-slice IR, enabling more aggressive noise suppression while preserving detail through multi-slice input. Training inputs are constructed by adding scaled noise-only patches (derived from the difference between 0.2 mm FBP and 0.2 mm IR) to 0.4 mm IR patches. Because 0.4 mm IR inherently has lower noise than 0.2 mm IR, a scaling factor of 1.0 would result in training inputs with overall noise lower than typical 0.2 mm FBP images used at inference. Scaling factors in the approximate range of 1.3–1.6 (varying depending on patient-specific noise statistics) are generally required to bring the training input noise level close to that observed across the 0.2 mm FBP images in our dataset. Random noise scaling (0.5–2.0) further broadens this range to cover clinical noise variability (e.g., due to patient size, CAREDose4D modulation), enhancing robustness and generalization. Random translations (−15 to +15 pixels) spatially decorrelate the inserted noise from anatomy and IR-specific patterns.²⁰ To empirically validate the contribution of thick-slice targets and noise scaling, we performed targeted ablation experiments. Results are summarized in Supplementary Figure A1. In contrast, projection-domain noise insertion remains the most accurate method for simulating realistic lower-dose noise before reconstruction,²⁵ but projection data are typically inaccessible from vendors. Pure image-domain alternatives, such as directly adding simulated noise to routine dose images to mimic lower-dose acquisitions,²⁶ are simpler but face challenges because image-domain noise is spatially variant and strongly influenced by reconstruction filters, beam hardening, and scatter. Many groups are actively working to improve image-domain noise simulation techniques to better approximate projection-domain realism and reduce these limitations.^27,28 Despite these trade-offs, our current method achieves substantial noise reduction (93% vs. FBP, 45% vs. IR) and excellent perceptual resolution gains (per radiologist scores), with no evidence of systematic IR artifacts in the denoised images.

Despite the advancements of our study, several limitations remain. First, the HR-CNN in this study was only trained for the Qr89 kernel. Reconstruction kernels sharper than Qr89 exist on the PCD-CT scanner. The reason why we focused on the Qr89 kernel was because it is the sharpest quantitative kernel without any edge enhancement on the system. For lung imaging, quantitative kernels are preferred. We can readily retrain the HR-CNN for other even sharper kernels for other clinical applications. Second, only a pilot clinical evaluation was performed. The sample size of five patient cases is relatively small. Future studies will include larger cohorts and more readers to improve statistical reliability. Finally, our evaluation relied on subjective image quality scores on clinical images. Future research will incorporate task-specific diagnostic performance metrics to better assess the clinical efficacy of the HR-CNN model.

5. Conclusion

Spatial resolution of PCD-CT is not fully utilized in routine practice. The proposed HR-CNN denoising method, along with small pixel size, may allow the high spatial resolution toward the system limit to be implemented in practice, which is beneficial in diagnosis of many diseases such as interstitial lung disease.