# Display Field-of-View Agnostic Robust CT Kernel Synthesis using Model-Based Deep Learning
Abstract
In X-ray computed tomography (CT) imaging, the choice of reconstruction kernel is crucial as it significantly impacts the quality of clinical images. Different kernels influence spatial resolution, image noise, and contrast in various ways. Clinical applications involving lung imaging often require images reconstructed with both soft and sharp kernels. The reconstruction of images with different kernels require raw sinogram data and storing images for all kernels increases processing time and storage requirements. The Display Field-of-View (DFOV) adds complexity to kernel synthesis, as data acquired at different DFOVs exhibit varying levels of sharpness and details. This work introduces an efficient, DFOV-agnostic solution for image-based kernel synthesis using model-based deep learning. The proposed method explicitly integrates CT kernel and DFOV characteristics into the forward model. Experimental results on clinical data, along with quantitative analysis of the estimated modulation transfer function using wire phantom data, clearly demonstrate the utility of the proposed method in real time. Additionally, a comparative study with a direct learning network, that lacks forward model information, shows that the proposed method is more robust to DFOV variations.
Index Terms — X-ray Computed Tomography, Kernel Synthesis, Model-based Deep Learning
1 Introduction
X-ray computed tomographic (CT) image Kernel Synthesis (KS) involves transforming an image reconstructed with one kernel into an image reconstructed with another kernel [1, 2, 3]. This process has various applications, including improving low contrast distinguishability, enhancing computer-aided detection [4], and facilitating quantitative analysis [5].
When reconstructing an image from raw sinogram data using the filtered back projection (FBP) algorithm, a reconstruction kernel is typically used to emphasize certain anatomical regions depending on the application. For instance, smooth kernels offer excellent low-contrast distinguishability with lower noise and fewer artifacts but have low spatial resolution. Conversely, sharp kernels provide high spatial resolution but also increase noise and artifacts. If raw sinogram data is unavailable or not stored due to large storage requirements, image-based kernel synthesis is used to reconstruct images with different kernels.
<details>
<summary>2502.14920v1/x1.png Details</summary>
### Visual Description
# Technical Document Extraction: Spectral Amplitude Analysis
## Chart 1: Input Spectral Amplitude
### Axes and Labels
- **X-axis**: `x Nyquist frequency (fN)`
- Markers: `0.0`, `0.2`, `0.4`, `0.6`, `0.8`, `1.0`
- **Y-axis**: `Input Spectral Amplitude`
- Range: `0.0` to `1.0` in increments of `0.2`
### Legend
- **Placement**: Right side of the chart
- **Labels and Colors**:
- `input MTF at DFOV 5 cm` (blue)
- `input MTF at DFOV 10 cm` (orange)
- `input MTF at DFOV 15 cm` (green)
- `input MTF at DFOV 20 cm` (red)
### Data Trends and Points
1. **Blue Line (5 cm DFOV)**:
- **Trend**: Steepest decline from `1.0` at `0.0 fN` to near `0.0` by `0.2 fN`.
- **Key Points**:
- `[0.0, 1.0]`
- `[0.2, ~0.0]`
2. **Orange Line (10 cm DFOV)**:
- **Trend**: Rapid decline, slower than blue.
- **Key Points**:
- `[0.0, 1.0]`
- `[0.4, ~0.0]`
3. **Green Line (15 cm DFOV)**:
- **Trend**: Gradual decline, flattest among input lines.
- **Key Points**:
- `[0.0, 1.0]`
- `[0.8, ~0.0]`
4. **Red Line (20 cm DFOV)**:
- **Trend**: Slowest decline, nearly flat after `0.6 fN`.
- **Key Points**:
- `[0.0, 1.0]`
- `[1.0, ~0.0]`
---
## Chart 2: Target Spectral Amplitude
### Axes and Labels
- **X-axis**: `x Nyquist frequency (fN)`
- Markers: `0.0`, `0.2`, `0.4`, `0.6`, `0.8`, `1.0`
- **Y-axis**: `Target Spectral Amplitude`
- Range: `0.0` to `1.75` in increments of `0.25`
### Legend
- **Placement**: Right side of the chart
- **Labels and Colors**:
- `Target MTF at DFOV 5 cm` (blue)
- `Target MTF at DFOV 10 cm` (orange)
- `Target MTF at DFOV 15 cm` (green)
- `Target MTF at DFOV 20 cm` (red)
### Data Trends and Points
1. **Blue Line (5 cm DFOV)**:
- **Trend**: Sharp peak at `0.1 fN`, then rapid decline.
- **Key Points**:
- `[0.0, 1.0]`
- `[0.1, ~1.75]`
- `[0.2, ~0.0]`
2. **Orange Line (10 cm DFOV)**:
- **Trend**: Broad peak centered at `0.15 fN`, higher amplitude than blue.
- **Key Points**:
- `[0.0, 1.0]`
- `[0.15, ~1.75]`
- `[0.3, ~0.0]`
3. **Green Line (15 cm DFOV)**:
- **Trend**: Lower peak at `0.2 fN`, slower decline.
- **Key Points**:
- `[0.0, 1.0]`
- `[0.2, ~1.5]`
- `[0.4, ~0.0]`
4. **Red Line (20 cm DFOV)**:
- **Trend**: Broadest peak at `0.25 fN`, lowest amplitude.
- **Key Points**:
- `[0.0, 1.0]`
- `[0.25, ~1.25]`
- `[0.5, ~0.0]`
---
## Cross-Reference Validation
- **Legend Consistency**:
- All colors in both charts match their respective DFOV labels (e.g., blue = 5 cm in both input and target charts).
- **Spatial Grounding**:
- Legends are positioned identically in both charts (right-aligned).
- **Trend Verification**:
- Input lines (Chart 1) show monotonic decay; target lines (Chart 2) exhibit resonant peaks inversely proportional to DFOV.
## Conclusion
The charts illustrate the relationship between Nyquist frequency and spectral amplitude for input and target MTF at varying DFOV distances. Input MTF amplitudes decay with increasing DFOV, while target MTF amplitudes exhibit resonant peaks that broaden and reduce in height with greater DFOV.
</details>
Fig. 1: Input (smooth) and target (sharp) kernel MTFs corresponding to different DFOVs. An increase in DFOV leads to high frequency content in the image.
However, image-based kernel synthesis remains a challenging inverse problem, as the synthesis must be robust to variations in the display field of view (DFOV) of the reconstructed images. For instance, Figure 1 illustrates the modulation transfer functions (MTFs) of smooth (input) and sharp (target) kernels at different DFOVs, ranging from 5 cm to 20 cm. It is evident from the plots that varying DFOVs result in different resolutions of reconstructed images. The effect of DFOV on image quality (IQ) is further demonstrated in Figure 2 using real water phantom data, where an increase in DFOV corresponds to increased speckle sharpness. In addtion to DFOV variations, managing artifacts and noise is another challenge in image-based kernel synthesis, especially when converting from smooth to sharp kernel images.
<details>
<summary>2502.14920v1/x2.png Details</summary>
### Visual Description
# Technical Document Analysis: DFOV Image Panels
## Image Description
The image consists of four grayscale panels labeled (a) through (d), each depicting a textured pattern with varying degrees of granularity and detail. The panels are arranged in a 2x2 grid, with labels positioned below each panel. No axes, legends, or numerical data tables are present in the image.
---
### Panel Analysis
#### (a) DFOV=5 cm
- **Label**: `(a) DFOV=5 cm`
- **Visual Characteristics**:
- Coarse, blurred texture with low spatial resolution.
- Large, indistinct regions of light/dark contrast.
- Minimal fine-scale detail visible.
- **Interpretation**: Represents the lowest Depth of Field of View (DFOV) setting, resulting in significant image blurring and loss of detail.
#### (b) DFOV=10 cm
- **Label**: `(b) DFOV=10 cm`
- **Visual Characteristics**:
- Moderate granularity with slightly improved clarity compared to Panel (a).
- Smaller textured elements visible, but still relatively coarse.
- Uniform distribution of light/dark patterns.
- **Interpretation**: Intermediate DFOV setting showing incremental improvement in detail retention.
#### (c) DFOV=15 cm
- **Label**: `(c) DFOV=15 cm`
- **Visual Characteristics**:
- Fine-grained texture with distinct small-scale features.
- Increased contrast between light/dark regions.
- Emergence of localized patterns (e.g., circular/elliptical shapes).
- **Interpretation**: Higher DFOV enables resolution of finer structural details, suggesting improved imaging sensitivity.
#### (d) DFOV=20 cm
- **Label**: `(d) DFOV=20 cm`
- **Visual Characteristics**:
- Highest spatial resolution among all panels.
- Dense, intricate patterns with sharp boundaries.
- Visible micro-scale variations in texture density.
- **Interpretation**: Maximum DFOV setting achieves optimal detail preservation, critical for high-resolution imaging applications.
---
### Key Observations
1. **DFOV-Resolution Relationship**:
- As DFOV increases (5 cm → 20 cm), spatial resolution improves exponentially.
- Blurring decreases, and fine-scale features become distinguishable.
2. **Texture Complexity**:
- Higher DFOV reveals increasingly complex microstructural patterns.
- Panel (d) exhibits fractal-like self-similarity at multiple scales.
3. **Contrast Dynamics**:
- Contrast uniformity increases with DFOV, enhancing feature delineation.
---
### Technical Implications
- **Imaging System Calibration**:
- DFOV settings directly impact diagnostic accuracy in imaging modalities (e.g., ultrasound, microscopy).
- Panel (d) would be optimal for applications requiring sub-millimeter detail.
- **Data Acquisition Tradeoffs**:
- Higher DFOV may require increased computational resources for processing.
- Panel (a) might be sufficient for gross anatomical studies but inadequate for cellular-level analysis.
---
### Limitations
- No color scale or quantitative metrics (e.g., pixel intensity values) are provided.
- Temporal or dynamic imaging data is absent; all panels represent static snapshots.
---
### Conclusion
This image demonstrates the critical role of DFOV in determining imaging resolution. Panels (a)-(d) provide a visual progression from severely blurred to highly detailed textures, emphasizing the importance of optimizing DFOV for specific imaging requirements.
</details>
Fig. 2: A $100× 100$ patch of a water phantom scanned at different DFOV values demonstrates significant texture changes as the DFOV increases. This dependency makes it challenging to develop a image-based kernel synthesis model that is agnostic to DFOV.
Ohkubo et al. in [6] proposed a direct kernel synthesis method that employs the ratio of modulation transfer functions (MTFs) between input and target kernels to transform images. This method is inherently robust to variations in the display field of view (DFOV) but often results in decreased image quality (IQ), particularly when converting smooth to sharp kernel images, due to amplified noise and artifacts in the reconstructed images.
Recently, several deep learning methods have been proposed to address kernel synthesis problem. Unlike direct kernel synthesis methods, these approaches are effective at managing noise during image transformation but struggle with variations in DFOVs. For example, a simple direct learning method like U-Net [7], trained on a specific DFOV, does not perform well when applied to data with different DFOVs. Similarly, experiments show that a network trained on data from all DFOVs tends to bias towards a specific DFOV.
To overcome this issue, we propose a model-based deep learning method that explicitly incorporates kernel MTFs into the deep learning framework. The proposed architecture for solving kernel synthesis problem consists of two parts: a deep learning-based projection step and an iterative data consistency solver that utilizes these MTFs. Enforcing data consistency during the training of the denoising network makes the kernel synthesis solution DFOV-agnostic, allowing a single network to be trained across various DFOVs with the advantage of requiring only a few hundred training samples.
This work complements our previous research [8] by developing a deep model in a supervised manner that is agnostic to DFOV using clinical data whereas previous work [8] focused on self-supervision using single-slice for noise aware kernel synthesis. A continuous kernel synthesis approach has been discussed in [9] but it is self-supervised. In contrast to [9], our proposed method is supervised learning with focused on DFOV and CT noise awareness.
2 Proposed Method
<details>
<summary>2502.14920v1/x3.png Details</summary>
### Visual Description
# Technical Document Extraction: Flowchart Analysis
## Diagram Description
The image depicts a **two-component iterative process** represented as a flowchart. The diagram uses rectangular blocks to denote computational steps and arrows to indicate data flow. All text is in **English**.
---
### Component Breakdown
#### 1. Neural Network based Projection Step
- **Label**: "Neural Network based Projection Step"
- **Input**: Variable `y` (arrow pointing from `y` to the block)
- **Output**: Variable `z` (arrow pointing from the block to `z`)
- **Spatial Position**: Leftmost block in the diagram
#### 2. DC Step with DFOV and Kernel Information
- **Label**: "DC Step with DFOV and Kernel Information, H"
- **Input**: Variable `z` (arrow pointing from `z` to the block)
- **Output**: Variable `x̂` (arrow pointing from the block to `x̂`)
- **Spatial Position**: Rightmost block in the diagram
#### 3. Iteration Mechanism
- **Label**: "Iterate"
- **Function**: Feedback loop from `x̂` back to `y`
- **Spatial Position**: Arrows connecting `x̂` → `y`
---
### Variable Flow
1. **Initial Input**: `y` → Neural Network Projection Step → `z`
2. **Intermediate Processing**: `z` → DC Step (with DFOV and Kernel Information) → `x̂`
3. **Feedback Loop**: `x̂` → Iterate → `y`
---
### Key Observations
- **No legends, axis titles, or numerical data** are present.
- **No secondary languages** detected; all text is in English.
- **No data tables or heatmaps** are included.
- **Spatial grounding**:
- `y` and `x̂` are positioned at the diagram's extremities.
- `z` acts as an intermediary variable between the two steps.
---
### Process Flow Summary
</details>
Fig. 3: The proposed training pipeline explicitly utilizes DFOV dependent water phantom data as noise together with input and target slice pairs. The network based project step acts as a generic DFOV agnostic denoiser that is shared across five unrolls used during network training. The analytical solution to the Data Consistency (DC) step is shown in Eq. (4). Incorporating the DC step explicitly into the learning framework helps in developing DFOV agnostic deep model.
The image-domain kernel synthesis can be expressed as a linear inverse problem of the form
$$
{\bm{y}}={\bm{\mathcal{H}}}{\bm{x}}+{\bm{n}}, \tag{1}
$$
where ${\bm{y}}∈\mathbb{R}^{N},N=p× q$ represents vectorized input kernel image of $p$ rows and $q$ columns, ${\bm{x}}∈\mathbb{R}^{N}$ is the target kernel image which needs to be synthesized, ${\bm{\mathcal{H}}}$ is the forward operator representing kernel synthesis process having input and target kernel information at specific DFOV. More specifically, using diagonalization, ${\bm{\mathcal{H}}}$ can be expressed as ${\bm{\mathcal{H}}}={\bm{\mathcal{F}}}^{T}\bm{\Lambda}{\bm{\mathcal{F}}}$ where, ${\bm{\mathcal{F}}}$ is the Fourier transform operator, $\bm{\Lambda}$ represents a DFOV dependent ratio of input and target CT kernel’s modulation transfer function (MTF). Unlike tradition inverse problems in imaging, ${\bm{n}}$ is a kernel and DFOV dependent noise that is assumed to be additive, as discussed in Fig. 2.
It is possible to utilize a direct learning approach such as UNet that takes ${\bm{y}}$ as input and ${\bm{x}}$ as target and learns a mapping $\hat{{\bm{x}}}=\text{CNN}({\bm{y}})$ between input and target pair of images to result in network prediction $\hat{{\bm{x}}}$ . However, such a direct learning approach depends on the large amount of training dataset consisting of DFOV variation which is difficult to collect for x-ray CT imaging. Further, without the explicit information about the Kernel’s MTF and DFOV, the learned model does not necessarily leads to data-consistent result as demonstrated in the experimental results in Fig. 4. *[Error downloading image: x4.png]*
Fig. 4: The top row shows inference results at a DFOV of 5 cm, and the bottom row at a DFOV of 10 cm. The proposed method in (c) produces sharper output compared to the direct learning method in (b). The green circle highlights artifacts (in zoomed version) in the DFOV 5 cm output of the direct learning method. The red arrow indicates hallucinations in the direct learning method, whereas the proposed method retains the structure well.
This work propose to find a solution to kernel synthesis inverse problem (1) using model-based deep learning [10] and represents the network regularized optimization problem in the form
$$
\arg\min_{{\bm{x}}{\bm{z}}}||{\bm{y}}-{\bm{\mathcal{H}}}{\bm{x}}||_{2}^{2}+%
\lambda||{\bm{x}}-{\bm{z}}||^{2}_{2}, \tag{2}
$$
where, ${\bm{z}}$ represents the output of a CNN that satisfy data consistency constraint as explicitly implied by the above problem formulation. Here, $\lambda$ is a regularization parameter initialized with $0.5$ and decayed with iteration, $\lambda_{k+1}=0.9\lambda_{k}$ , to gradually give more importance to data consistency term. The above problem (2) can be solved iteratively in two steps using alternating minimization as follows:
$$
\displaystyle{\bm{z}}_{k} \displaystyle=\text{CNN}({\bm{x}}_{k}) \displaystyle{\bm{x}}_{k+1} \displaystyle={\bm{\mathcal{F}}}^{T}\left(\frac{{\bm{\mathcal{F}}}({\bm{%
\mathcal{H}}}^{T}{\bm{y}}+\lambda{\bm{z}}_{k})}{|\bm{\Lambda}|^{2}+\lambda}%
\right), \tag{3}
$$
where ${\bm{x}}_{0}$ can be initialized as Tikhonov solution of the form ${\bm{x}}_{0}=({\bm{\mathcal{H}}}^{T}{\bm{\mathcal{H}}}+\lambda\bm{\mathcal{I}}%
)^{-1}{\bm{\mathcal{H}}}^{T}{\bm{y}}$ , here $\bm{\mathcal{I}}$ is identity matrix. The analytical solution in Eq. (4) assumes ${\bm{z}}_{k}$ to be constant for the data-consistency problem.
As visually demonstrated in Fig. 2, DFOV impacts the texture of water phantom data, which is used as noise during network training. Fig. 3 presents a schematic diagram of the end-to-end network training strategy using supervised learning with existing clinical data from different DFOVs. The use of DFOV-dependent water phantom data as noise allows for controlling the enhancement of noise and artifacts during the iterative reconstruction process.
3 Trainings and Experiments
Two separate supervised training setups were used. The first was a direct learning method that trained a U-Net to perform combined DFOV training without kernel MTF information. The second was the proposed method, which employed a U-Net with weight sharing across five unrolls and incorporated explicit DFOV and kernel information in the forward model.
The networks were trained using existing clinical lung images obtained from the GE Revolution Ascend system, acquired with 120 kVp and 544 mA, and reconstructed with different DFOV values of 5 cm, 10 cm, 15 cm, and 20 cm. The training data consisted of 1280 slices, each of size $512× 512$ , from different subjects. STANDARD kernel images were used as input, and LUNG kernel images were used as target images for all DFOVs. Both networks were trained for 500 epochs with an initial learning rate of $1× 10^{-4}$ , using mean square error (MSE) combined with the structural similarity index (SSIM) as the loss function. Following the training, the performance of the trained networks was verified using both clinical and phantom data across various DFOV.
In the first experiment, we compared the reconstruction quality of the proposed method with that of the direct learning method. Fig. 4 shows portions of the input and target kernel images, along with the outputs of the direct learning method and the proposed method for DFOV 5 cm and DFOV 10 cm, respectively, in the top and bottom rows. The proposed method achieves sharpness closer to the ground truth while preserving structural information, unlike the direct learning method, which results in blurred images with hallucinations, as indicated by the green circle and red arrow.
In the second experiment, we quantitatively evaluated the performance of the proposed method by estimating the MTF on the DFOV 10 cm wire phantom data. Figure 5 (a) shows the wire phantom images for the input, target, direct learning method, and proposed method. It is evident from the images that the wire image produced by the proposed method is closer to the target wire image compared to the wire image from the direct learning method. Figure 5 (b) displays the MTFs estimated for the proposed method and the direct learning method, compared with the input and target MTFs. The estimated MTF by the proposed method shows a higher mid-frequency boost compared to the direct learning method and is closer to the target.
<details>
<summary>2502.14920v1/x5.png Details</summary>
### Visual Description
# Technical Document Analysis of Image Panels
## Panel Descriptions
### Panel (a) Input
- **Label**: "(a) Input" (bottom-left corner)
- **Visual Characteristics**:
- Grayscale image with low contrast
- Single bright white spot centered in the frame
- Background texture appears granular/noisy
- No discernible patterns or structures
### Panel (b) Direct
- **Label**: "(b) Direct" (bottom-left corner)
- **Visual Characteristics**:
- Similar grayscale background to Panel (a)
- Bright white spot with a distinct dark halo/rim
- Increased contrast around the central spot
- Background noise patterns more pronounced
### Panel (c) Proposed
- **Label**: "(c) Proposed" (bottom-left corner)
- **Visual Characteristics**:
- High-contrast grayscale image
- Central bright spot with sharp, defined edges
- Background features enhanced with localized intensity variations
- Appears to emphasize spatial resolution improvements
### Panel (d) Target
- **Label**: "(d) Target" (bottom-left corner)
- **Visual Characteristics**:
- Highest contrast among all panels
- Central bright spot with perfect circular symmetry
- Background features exhibit structured, repeating patterns
- Likely represents idealized/ground truth reference
## Observations
1. **Progression**: Panels progress from raw input (a) to processed outputs (b,c) toward a target reference (d)
2. **Contrast Enhancement**: Each subsequent panel shows increased contrast and structural definition
3. **Spot Characteristics**:
- Panel (a): Diffuse, low-intensity spot
- Panel (b): Moderate-intensity spot with halo
- Panel (c): High-intensity spot with defined edges
- Panel (d): Idealized spot with perfect symmetry
## Structural Analysis
- **Layout**: 2x2 grid format with equal panel sizing
- **Labeling Convention**: Standardized "(letter) Label" format in bottom-left corner
- **Color Scheme**: Monochromatic grayscale throughout all panels
- **Absent Elements**: No legends, axes, or numerical data present
## Conclusion
The image presents a comparative visualization of image processing stages, with Panel (d) Target serving as the reference standard. The sequence demonstrates increasing image quality and feature definition from raw input through proposed processing methods.
</details>
(a) Wire phantom
<details>
<summary>2502.14920v1/x6.png Details</summary>
### Visual Description
# Technical Document Extraction: Line Graph Analysis
## Image Description
The image is a line graph comparing four data series across a logarithmic scale. The graph features four distinct colored lines representing different datasets, with a legend positioned in the top-right corner outside the plot area.
## Axis Labels and Markers
- **X-axis**: "Line pairs per millimeter" (logarithmic scale)
- Range: 0.0 to 2.5
- Increment: 0.5 units
- **Y-axis**: "Amplitude"
- Range: 0.00 to 1.75
- Increment: 0.25 units
## Legend
- **Position**: Top-right corner (outside plot area)
- **Entries**:
1. **Blue**: Input
2. **Orange**: Direct Learning
3. **Green**: Proposed Network
4. **Red**: Target
## Data Series Analysis
### 1. Input (Blue Line)
- **Trend**: Monotonically decreasing from 1.0 amplitude at x=0.0 to 0.0 at x=1.0
- **Key Points**:
- [0.0, 1.0]
- [0.5, 0.25]
- [1.0, 0.0]
### 2. Direct Learning (Orange Line)
- **Trend**: Initial rise to 1.25 amplitude at x≈0.3, then sharp decline to 0.0 at x=1.0
- **Key Points**:
- [0.0, 1.0]
- [0.3, 1.25]
- [1.0, 0.0]
### 3. Proposed Network (Green Line)
- **Trend**: Slightly higher peak than Input (1.5 amplitude at x≈0.4), then rapid decline to 0.0 at x=1.0
- **Key Points**:
- [0.0, 1.0]
- [0.4, 1.5]
- [1.0, 0.0]
### 4. Target (Red Line)
- **Trend**: Highest peak (1.75 amplitude at x≈0.3), then steep decline to 0.0 at x=1.0
- **Key Points**:
- [0.0, 1.0]
- [0.3, 1.75]
- [1.0, 0.0]
## Post-X=1.0 Behavior
All data series converge to 0 amplitude and remain flat at y=0 for x > 1.0.
## Validation Checks
1. **Color Consistency**: All legend colors match corresponding lines in the plot
2. **Trend Verification**:
- Input shows strict monotonic decrease
- Direct Learning and Proposed Network exhibit single-peak behavior
- Target demonstrates highest peak amplitude
3. **Axis Alignment**: All data points align with logarithmic x-axis scaling
## Conclusion
The graph illustrates comparative performance metrics across four datasets, with the Target dataset demonstrating the highest initial amplitude and fastest convergence to baseline. The Proposed Network shows intermediate performance between Input and Target datasets.
</details>
(b) Estimated MTFs at DFOV 10 cm
Fig. 5: (a) Estimated wire phantom images visually show that proposed method results in sharper output compared to a direct learning method. (b) Quantitative results at DFOV 10 cm on wire phantom data demonstrate that the proposed method produces sharper output compared to the direct learning method. The orange curve represents the direct learning method, which has less mid-frequency boost compared to the green curve representing the proposed method.
4 Conclusions
This work presented a DFOV-agnostic image-based kernel synthesis method that performs kernel synthesis which is robust to input noise. This method significantly improves disease diagnosis and treatment by enhancing the visibility and precision of anatomical structures such as bones, vessels, and lesions as shown by converting smooth kernel images to sharp kernel images.
5 Compliance with ethical standards
This research study was conducted using human subject data made available through appropriate research contracts. All the activities necessary for obtaining the research results/development were carried out with due approval from the ethics committee and competent authorities in compliance with the regional laws and regulations.
References
- [1] Jooae Choe, Sang Min Lee, Kyung-Hyun Do, Gaeun Lee, June-Goo Lee, Sang Min Lee, and Joon Beom Seo, “Deep Learning–based Image Conversion of CT Reconstruction Kernels Improves Radiomics Reproducibility for Pulmonary Nodules or Masses,” Radiology, vol. 292, no. 2, pp. 365–373, 2019.
- [2] Jakob Neubauer, Eva Maria Spira, Juliane Strube, Mathias Langer, Christian Voss, and Elmar Kotter, “Image quality of mixed convolution kernel in thoracic computed tomography,” Medicine, vol. 95, no. 44, 2016.
- [3] Masaki Ohkubo, Akihiro Narita, Shinichi Wada, Kohei Murao, and Toru Matsumoto, “Image filtering to make computer-aided detection robust to image reconstruction kernel choice in lung cancer CT screening,” Medical Physics, vol. 43, no. 7, pp. 4098–4105, 2016.
- [4] Tianzhong Lan, Fanxin Zeng, Zhang Yi, Xiuyuan Xu, and Min Zhu, “Icnodulenet: Enhancing pulmonary nodule detection performance on sharp kernel ct imaging,” IEEE Journal of Biomedical and Health Informatics, 2024.
- [5] Naoya Tanabe, Shizuo Kaji, Hiroshi Shima, Yusuke Shiraishi, Tomoki Maetani, Tsuyoshi Oguma, Susumu Sato, and Toyohiro Hirai, “Kernel conversion for robust quantitative measurements of archived chest computed tomography using deep learning-based image-to-image translation,” Frontiers in Artificial Intelligence, vol. 4, pp. 769557, 2022.
- [6] Masaki Ohkubo, Shinichi Wada, Akihiro Kayugawa, Toru Matsumoto, and Kohei Murao, “Image filtering as an alternative to the application of a different reconstruction kernel in CT imaging: feasibility study in lung cancer screening,” Medical Physics, vol. 38, no. 7, pp. 3915–3923, 2011.
- [7] Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-Net: Convolutional networks for biomedical image segmentation,” in Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. Springer, 2015, pp. 234–241.
- [8] Hemant Kumar Aggarwal, Phaneendra K Yalavarthy, and Rajesh Langoju, “Self-supervised noise-aware kernel synthesis for improved x-ray computed tomography imaging,” in 2024 IEEE International Symposium on Biomedical Imaging (ISBI). IEEE, 2024, pp. 1–4.
- [9] Serin Yang, Eung Yeop Kim, and Jong Chul Ye, “Continuous conversion of ct kernel using switchable cyclegan with adain,” IEEE Transactions on Medical Imaging, vol. 40, no. 11, pp. 3015–3029, 2021.
- [10] Hemant K Aggarwal, Merry P Mani, and Mathews Jacob, “MoDL: Model-based deep learning architecture for inverse problems,” IEEE Transactions on Medical Imaging, vol. 38, no. 2, pp. 394–405, 2018.