CT Image Quality Assessment Method Based on Poisson Flow Generation Model

Abstract

This invention discloses a CT image quality assessment method based on a Poisson flow generation model. A pseudo-reference image is generated by a pseudo-reference image generation module (PPFM) to compensate for the lack of a reference image and help the subsequent score regression model capture detailed features. A hierarchical feature fusion regression module fuses upper-layer and lower-layer features layer by layer to make full use of features at different scales. At the same time, the features extracted in the last layer are sequentially passed to ViT (Vision Transformer module) to fuse the obtained global and local features together to predict the final image quality.

Description

Technical Field

[0001] This invention pertains to medical imaging and computer vision technologies, specifically relating to a CT image quality assessment method based on a Poisson flow generation model. Background Technology

[0002] Computed tomography (CT) is one of the most commonly used medical imaging techniques, playing an indispensable role in doctors' diagnostic processes. However, the high doses of radiation during CT scans can cause harm to patients, posing significant health risks. Therefore, imaging with lower radiation doses has become an important technique for reducing radiation exposure risks. However, while low-dose CT radiation reduces radiation, it also decreases image quality, introducing noise and artifacts that can interfere with accurate diagnosis and affect subsequent image processing tasks such as image segmentation, restoration, and enhancement. Therefore, advancing precise image quality assessment (IQA) methods is crucial for optimizing CT scan radiation doses, ensuring patient safety, and ensuring the reliability of diagnostic results.

[0003] For a given image, the goal of Image Quality Assurance (IQA) is to predict a score as close as possible to the Human Mean Subjective Rating (MOS) using image quality assessment algorithms. Depending on whether a reference image is required, IQA can be categorized into full-reference image quality assessment, partial-reference image quality assessment, and no-reference image quality assessment. Full-reference image quality assessment uses information from both the reference and distorted images to calculate the quality score, while partial-reference image quality assessment uses information from both the distorted and partial reference images. Full-reference and partial-reference image quality assessments exhibit good performance due to their reliance on a reference image to calculate the image quality score. Therefore, Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) are widely used as standard metrics for image quality assessment because they are computationally efficient and strongly correlated with human visual perception of traditional distortion types such as blur, compression, and Gaussian noise. However, in real-world scenarios, reference images are often unavailable, limiting the applicability of these methods in practice. This problem is particularly relevant in the medical field, where obtaining high-quality reference images can have adverse effects on patients. Therefore, the no-reference image quality assessment method does not require a reference image to calculate the image score, making it more suitable for practical applications and worthy of further development.

[0004] Traditional no-reference image quality assessment (NR-IQA) methods primarily rely on the extraction of handcrafted features to evaluate image quality by simulating the perceptual characteristics of the human visual system. After image preprocessing, various handcrafted features are extracted, including statistical features (such as mean and variance), texture features (such as Gabor filters), edge features (such as the Sobel operator), structural features (such as local binary patterns), and frequency domain features (such as Fourier transforms). These features are fused using different methods to extract their statistical distribution. Finally, a mapping function, such as support vector regression, is designed to predict the quality of the input image. However, due to the complexity of image content and distortion patterns, the representational power of handcrafted features is often unsatisfactory.

[0005] Recently, Convolutional Neural Networks (CNNs) have attracted widespread attention in the NR-IQA task due to their powerful capabilities. Hyper-IQA categorizes features into low-level and high-level classes, then transforms the latter to reshape the influence of the former. SAHN developed an adaptive super network to aggregate local distortion features and global semantic features. Meta-IQA proposed a meta-learning framework for IQA to improve generalization to unknown image distortions. Other methods add attention mechanisms to CNNs to enrich local fine-grained details and multi-scale global semantic information. MANIQA proposed a multi-dimensional attention network for no-reference image quality assessment. MUSIQ proposed a multi-scale image quality Transformer, utilizing the Transformer architecture to address the problem of images with different sizes and aspect ratios.

[0006] Meanwhile, many researchers have also explored the field of medical image quality assessment. Due to the diverse characteristics of medical imaging, many different methods exist for specific imaging modalities. These include medical image types such as MRI, ultrasound imaging, and X-rays. There has also been considerable research in the field of CT image quality assessment. SSIQA proposes a no-reference IQA model for CT images by jointly learning noise level estimation and image quality quantification. MD-IQA introduces semi-supervised learning and uses pseudo-labels from unlabeled data to guide model training. KFKT-IQA uses natural images from similar tasks for transfer learning. Other methods use task-based IQA models to detect virtual inserted objects with simple geometries and calculate the quantitative quality of CT images.

[0007] However, the lack of reference information presents a challenge, making these methods consistently bottlenecked compared to FR-IQA and potentially negatively impacting their overall performance. Therefore, some recent methods employ reconstruction-based strategies to learn quality-aware features from the original image. In this context, frameworks based on Generative Adversarial Networks (GANs) have been proposed. For example, Lin et al. developed a quality-aware GAN to generate illusory references conditioned on distorted images. Similarly, Ma et al. proposed generating the main content of distorted images. In another study, Ren et al. proposed a restorative adversarial network to reconstruct the input distorted image. Recently, D-BIQA proposed using a conditional denoising diffusion model to generate the main content of low-quality images. However, the generative models of these methods are more suitable for some quality degradation problems in natural images and less applicable to CT images, because noise in CT images is generally quantum noise that follows Poisson noise, i.e., random fluctuations of photons cause granular noise in the image. Summary of the Invention

[0008] Purpose of the Invention: The purpose of this invention is to address the shortcomings of existing technologies and provide a CT image quality assessment method based on the Poisson flow generation model. A pseudo-reference image is generated by the pseudo-reference image generation module PPFM to compensate for the lack of a reference image and help the subsequent score regression model capture detailed features. The hierarchical feature fusion regression module fuses upper-layer and lower-layer features layer by layer to fully utilize features at different scales. Simultaneously, the features extracted from the last layer are sequentially passed to the ViT (Vision Transformer module) to fuse the obtained global and local features together to predict the final image quality.

[0009] Technical solution: The present invention provides a CT image quality assessment method based on a Poisson flow generation model, comprising the following steps:

[0010] Step 1: Construct an image quality assessment network, which includes a pseudo-reference image generation module (PPFM) and a hierarchical feature fusion regression module; the specific details are as follows:

[0011] Step 1.1: The pseudo-reference image generation module PPFM, based on the Poisson flow generation model, generates the original low-dose CT image I... d The pseudo-reference image G is obtained after denoising by the input pseudo-reference image generation module PPFM. θ (I d );

[0012] Step 1.2: Transfer the original low-dose CT images I d And the corresponding difference diagram I e and dissimilar graph I dis The images are fused to obtain a three-channel image. H and W represent the height and width of the image, respectively;

[0013] Step 1.3: Input the three-channel image E into the hierarchical feature fusion regression module. The hierarchical feature fusion regression module includes a trapezoidal CNN network. The trapezoidal CNN network has a total of N layers. Each layer has a corresponding number of bottleneck blocks. A Vision Transformer module is added after each bottleneck block in the last layer of the CNN network.

[0014] After the three-channel image E enters the CNN network, the feature map F extracted from the i-th layer is obtained. i Feature alignment is performed using bottleneck blocks to obtain aligned feature maps. Alignment feature map With the feature map F of the next layer i+1 Then, the bottleneck block is used for fusion and alignment. The alignment features of the first N-1 layers are processed by the Vision Transformer module to obtain the features. The feature map F extracted from the last layer of the CNN network N With features The feature F is obtained by fusion, and the feature F is passed through a pooling layer to obtain the prediction quality score;

[0015] Step 2: Construct the loss function L total Optimize the image quality assessment network, loss function L total Including MSE loss L quality Image relative ranking loss L relative .

[0016] Furthermore, the difference diagram I e and dissimilar graph I dis The calculation method is as follows:

[0017] I e =|I d -G θ (I d )|

[0018] I dis =(1-SSIM(I) d G θ (I d )))

[0019] In the above formula, θ represents the network parameters, and SSIM() is the structural similarity measurement function. Here, SSIM() is used to measure the similarity between two images.

[0020] Since the number of channels and dimensions vary in each layer of the CNN network, a bottleneck block consisting of three convolutional operations is set up to integrate the features extracted from each layer into the final feature representation. This bottleneck block includes a 1×1 convolutional layer, a 3×3 convolutional layer, and a 1×1 convolutional layer, respectively, to align the features. The calculation formula is as follows:

[0021]

[0022] Among them, W 1×1 and W 3×3 These represent the weight matrices for the 1×1 and 3×3 convolutional layers, respectively; thus, the feature map F... i The number of channels and resolution were adjusted to an appropriate level;

[0023] The number of bottleneck blocks in the i-th layer of the CNN network is i;

[0024] The formula for calculating feature F is as follows:

[0025]

[0026] Furthermore, the loss function L total The calculation formula is as follows:

[0027] L total =L quality +L relative

[0028] The MSE loss is calculated as follows:

[0029]

[0030] In the above formula, B represents the total number of images in each batch, and s k and Let represent the true quality score (subjective quality score) and the predicted quality score of the k-th image, respectively.

[0031] The relative ranking loss of the image is calculated as follows:

[0032] L relative =L triplet (q max ,q′ max ,q min )+L rriplet (q min ,q′ min ,q max )

[0033] =max{0,d(q′) max ,q max )-d(q max ,q min)+margin1}+max{0,d(q′ min ,q min )-d(q max ,q min +margin2}

[0034] Within each batch of CT images, q max ,q′ max ,q min ,q′ min These represent the predicted quality scores of the images with the highest, second-highest, lowest, and second-lowest subjective quality scores, respectively.

[0035] Let q represent the predicted quality score. max The subjective quality scores corresponding to the images are represented by these images, and the representation rules for the remaining images are similar.

[0036] set up as well as

[0037] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0038] (1) The present invention generates corresponding pseudo-reference images based on the Poisson flow generation model. The smaller the difference between the pseudo-reference image and the real reference image, the more accurate the judgment of the quality regression network. This helps the subsequent fractional regression model capture detailed features.

[0039] (2) The hierarchical feature fusion regression network of the present invention includes ViT and CNN modules to fuse global features with local features to predict the final image quality. The CNN module has an overall trapezoidal structure, hierarchically integrating features extracted from the middle layer, enabling the model to make full use of visual information features from low to high layers;

[0040] (3) Since artifacts in CT images are radial, this invention adds a Vision Transformer after the last layer of CNN feature extraction to fuse the acquired global features with local features.

[0041] (4) This invention introduces relative ranking loss to construct the total loss function in order to make full use of the relative relationship between data. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the overall network structure of the present invention;

[0043] Figure 2 A schematic diagram of the PPFM workflow for mapping low-dose CT images to normal images;

[0044] Figure 3 This is a schematic diagram of the workflow for obtaining three-channel images by processing low-dose CT images and pseudo-reference images generated by PPFM.

[0045] Figure 4 This is a comparison image of the various technical solutions in the embodiments. Detailed Implementation

[0046] The technical solution of the present invention will be described in detail below, but the scope of protection of the present invention is not limited to the embodiments described.

[0047] This invention's CT image quality assessment network comprises two main parts: a pseudo-reference image generation module (PPFM) based on a Poisson flow generation model, and a hierarchical feature fusion regression module combining CNN and Transformer. Research indicates that two main factors influence CT image quality: noise and artifacts, with quantum noise following a Poisson distribution. Therefore, this invention uses a Poisson flow generation model to generate corresponding pseudo-reference images to compensate for missing reference images and help the subsequent score regression model capture detailed features. Experiments show that by adding pseudo-reference images, this invention improves model performance across all evaluation metrics. The hierarchical feature fusion regression module employs a stepped structure to fuse upper-layer and lower-layer features layer by layer, fully utilizing features at different scales. Simultaneously, ViT is used to fuse the obtained global and local features to predict the final image quality.

[0048] like Figure 1 As shown, the CT image quality assessment method based on the Poisson flow generation model of the present invention includes the following steps:

[0049] Step 1: Construct an image quality assessment network, which includes a pseudo-reference image generation module (PPFM) and a hierarchical feature fusion regression module; the specific details are as follows:

[0050] Step 1.1: The pseudo-reference image generation module PPFM, based on the Poisson flow generation model, generates the original low-dose CT image I... d The pseudo-reference image G is obtained after denoising by the input pseudo-reference image generation module PPFM. θ (I d );

[0051] Step 1.2: Transfer the original low-dose CT images I d And the corresponding difference diagram I e and dissimilar graph I dis The images are fused to obtain a three-channel image. H and W represent the height and width of the image, respectively;

[0052] Step 1.3: Input the three-channel image E into the hierarchical feature fusion regression module. The hierarchical feature fusion regression module includes a trapezoidal CNN network. The trapezoidal CNN network has a total of N layers. Each layer has a corresponding number of bottleneck blocks. A Vision Transformer module is added after each bottleneck block in the last layer of the CNN network.

[0053] After the three-channel image E enters the CNN network, the feature map F extracted from the i-th layer is obtained. i Feature alignment is performed using bottleneck blocks to obtain aligned feature maps. Alignment feature map With the feature map F of the next layer i+1 Then, the bottleneck block is used for fusion and alignment. The alignment features of the first N-1 layers are processed by the Vision Transformer module to obtain the features. The feature map F extracted from the last layer of the CNN network N With features The feature F is obtained by fusion, and the feature F is passed through a pooling layer to obtain the prediction quality score;

[0054] Step 2: Construct the loss function L total Optimize the image quality assessment network, loss function L total Including MSE loss L quality Image relative ranking loss L relative .

[0055] Furthermore, the difference diagram I e and dissimilar graph I dis The calculation method is as follows:

[0056] I e =|I d -G θ (I d )|

[0057] I dis =(1-SSIM(I) d G θ (I d )))

[0058] In the above formula, θ represents the network parameters.

[0059] In this embodiment, the bottleneck block sequentially includes a 1×1 convolutional layer, a 3×3 convolutional layer, and a 1×1 convolutional layer, aligning features. The calculation formula is as follows:

[0060]

[0061] Among them, W 1×1 and W3×3 These represent the weight matrices for 1×1 and 3×3 convolutional layers, respectively.

[0062] The number of bottleneck blocks in the i-th layer of the CNN network is i;

[0063] The formula for calculating feature F is as follows:

[0064] To further improve the feature modeling capability of the image quality assessment network, this embodiment introduces a corresponding number of Vision Transformer modules into the network. Each Vision Transformer module includes components such as position encoding, an attention mechanism, and a multilayer perceptron. The attention mechanism is calculated as follows:

[0065]

[0066] The projections of the query (Q), key (K), and value (V) are achieved through three independent linear projections. k Represents the spatial dimensions of Q, K, and V.

[0067] The original low-dose CT images were processed using PPFM (e.g.) Figure 2 (As shown) Generates a corresponding high-quality pseudo-reference image, and the processed multi-channel image is input into a hierarchical feature fusion regression network (such as...). Figure 3 Feature extraction is performed in the image (as shown). First, the multi-channel image is input into the preprocessing backbone network ResNet50. The features extracted from each layer are aligned through the bottleneck network, and then the features from different layers are fused layer by layer. After five layers of feature extraction, the last layer of features is input into VIT, where the extracted local features are fused with the global features. Finally, the resulting features are processed through pooling layers to obtain the quality score predicted by the image quality assessment network.

[0068] To utilize the features extracted by the convolutional network, this embodiment uses a fully connected layer as a fusion layer to map the aforementioned features and predict the perceptual quality of the image. The final loss function L... total The calculation formula is as follows:

[0069] L total =L quality +L relative

[0070] The MSE loss is calculated as follows:

[0071]

[0072] In the above formula, B represents the total number of images in each batch, and s k and Let represent the true quality score (subjective quality score) and the predicted quality score of the k-th image, respectively.

[0073] While MSE loss is effective for quality prediction tasks, it does not explicitly consider the ranking and correlation between images. Since the goal of this invention is to consider the relative ranking relationship between each batch of samples, this embodiment introduces image relative ranking loss.

[0074] The relative ranking loss of the image is calculated as follows:

[0075] L relative =L triplet (q max ,q′ max ,q min )+L triplet (q min ,q′ min ,q max )

[0076] =max{0,d(q′) max ,q max )-d(q max ,q min )+margin1}+max{0,d(q′ min ,q min )-d(q max ,q min +margin2}

[0077] Within each batch of CT images, q max ,q′ max ,q min ,q′ min These represent the predicted quality scores of the images with the highest, second-highest, lowest, and second-lowest subjective quality scores, respectively.

[0078] Let q represent the predicted quality score. max The subjective quality score corresponds to the image shown; the representation rules for the remaining images are similar. as well as

[0079] Example 1

[0080] This embodiment was implemented on an NVIDIA GeForce RTX 3090 equipped with PyTorch 1.12.1 and CUDA 11.3 for both training and testing. To train the Poisson flow generation model, the image quality assessment network was trained for 100k iterations with D=64 and a learning rate of 2×10⁻⁶. -4The batch size is 32. After preprocessing, exponential moving average (EMA) planning, and leak-free enhancement, the enhancement probability is 15%. In this embodiment, the discard probability is set to 10%.

[0081] The training dataset used was the Mayo Low-Dose CT dataset, a public clinical dataset containing CT images of 10 patients, without human-perceived scores for each image. The data was split into a training set (4800 slices) containing the first 8 patients and a validation set (1136 slices) containing the last 2 patients. After training, corresponding pseudo-reference images were obtained from the LDCTBIQA2023 challenge dataset using the trained PPFM. The LDCTBIQA2023 challenge dataset contains a total of 1000 distorted abdominal images with noise and streak artifacts, generated at four different dose levels: 100%, 50%, 25%, and 10%. Of these images, 900 were allocated to the training phase, and the remaining 100 were reserved for the testing phase. The final human-perceived score for each image was determined by averaging individual scores from five skilled radiologists. Subsequently, corresponding difference maps and dissimilarity maps were calculated and... Figure 3 Multi-channel images are obtained by connecting them in this way.

[0082] The obtained multi-channel images are used as input to a hierarchical feature fusion regression network. First, they are fed into a preprocessed ResNet50 network to obtain features F from different layers. i In this embodiment, N is set to 5, meaning the CNN network in this embodiment has five layers; then, through the bottleneck network, 1×1, 3×3, and 1×1 convolutional layers are used sequentially, and the feature map F i The number of channels and resolution were adjusted to Then The features are sequentially fused with those of the next layer to fully learn the features from different layers. Finally, the features from the 5th layer are input into ViT to obtain global features, which are then passed through a pooling layer to obtain the predicted score of this invention. During training, the learning rate is set to 4×10. -5 The batch size B is set to 16. The StepLR scheduler is used with a step size of 25 periods and a decay factor of 0.7. The model is regularized using an exponential moving average (EMA) with a decay rate of 0.997.

[0083] Example 2

[0084] This embodiment selects eight NR-IQA methods for performance comparison, using Pearson linear correlation coefficient (PLCC), Spearman rank correlation coefficient (SROCC), Kendall rank correlation coefficient (KROCC), and their sum (Overall) as metrics to evaluate model performance. The comparison methods include CNN-based methods (DBCNN, HyperIQA, and QPT), Transformer-based methods (TReS, AHIQ, MANIQA, and MD-IQA), and the D-BIQA method based on a denoising diffusion probability model.

[0085] Table 1 compares the results with other methods on the LDCTIQA2023 dataset.

[0086] PLCC SROCC KROCC OverAll DBCNN 0.9716 0.9693 0.8692 2.8101 HyperIQA 0.9680 0.9694 0.8672 2.8046 QPT 0.9743 0.9732 0.8797 2.8272 TReS 0.9755 0.9745 0.8786 2.8286 AHIQ 0.9762 0.9746 0.8810 2.8318 MANIQA 0.9768 0.9786 0.8891 2.8445 MD-IQA 0.9789 0.9792 0.9041 2.8622 D-BIQA 0.9814 0.9816 0.9122 2.8752 This invention 0.9815 0.9837 0.9113 2.8765

[0087] As shown in Table 1, the technical solution of this invention outperforms state-of-the-art methods in SROCC, PLCC, and overall evaluation on the LDCTIQA2023 database. Compared with three CNN-based methods, the image quality assessment network of this invention improves the overall score by at least 5%. Compared with other methods that add Transformers, the overall score is also improved by 1% to 5%. It should be noted that compared with D-BIQA, which also uses a generative model to obtain corresponding high-quality images, the technical solution of this invention outperforms D-BIQA in all three metrics. Furthermore, the pseudo-reference image generation speed of PPFM in the technical solution of this invention is nearly 490 times faster than that of general diffusion models, as shown in Table 2.

[0088] Table 2 Ablation Study of Generative Models

[0089] Generative Models Overall AIT diffusion model 2.7767 94.531 This invention PPFM 2.8765 0.192

[0090] Example 3

[0091] This embodiment uses the technical solution of the present invention to visualize the CT images and thermal images of five representative examples in the LDCTBIQA2023 test dataset, such as... Figure 4 As shown, Figure 4 From top to bottom are low-dose CT images, thermal images, MOS (GT), and the score predicted by this invention. Figure 4 The heatmap clearly highlights key elements that strongly influence human perception. When humans view an image, features of interest significantly impact their perception; therefore, these features have a greater influence on the image quality assessment network of this invention.

[0092] To evaluate the efficiency of each module in the network of this invention, ablation experiments were conducted. A pre-trained ResNet50 was used as the backbone model of the image quality assessment network, and the performance of each individual component was analyzed by comparing SROCC, KROCC, and Overall.

[0093] Table 3 Ablation studies of each component

[0094] PPFM HFFRN Lr SROCC KROCC OverAll 0.9752 0.8914 2.8439 √ 0.9794 0.9051 2.8646 √ 0.9816 0.9043 2.8680 √ √ 0.9816 0.9089 2.8719 √ √ √ 0.9837 0.9113 2.8765

[0095] As shown in Table 3, this embodiment first adds a posterior sampling Poisson flow generation model (PPFM), and processes the pseudo-reference map generated by PPFM as the network input. This results in improvements in SROCC and KROCC metrics, increasing by 0.6% and 1.3%, respectively. Then, the effectiveness is examined by connecting the proposed hierarchical feature fusion regression network (HFFRN) with the ResNet50 output features. SROCC and KROCC show significant improvements of approximately 0.4% and 1.4%, respectively. Performance is further improved when PPFM and HFFRN are added simultaneously. Finally, the relative ranking loss (Lr) is added, further boosting SROCC and KROCC to maximum values ​​of 98.37% and 91.13%, respectively, achieving an overall SROCC of 2.8765.

Claims

1. A CT image quality assessment method based on a Poisson flow generation model, characterized in that, Includes the following steps: Step 1: Construct an image quality assessment network, which includes a pseudo-reference image generation module (PPFM) and a hierarchical feature fusion regression module; the specific details are as follows: Step 1.1: The pseudo-reference image generation module PPFM, based on the Poisson flow generation model, generates the original low-dose CT image I... d The pseudo-reference image G is obtained after denoising by the input pseudo-reference image generation module PPFM. θ (I d ); Step 1.2: Transfer the original low-dose CT images I d And the corresponding difference diagram I e and dissimilar graph I dis The images are fused to obtain a three-channel image E. H and w represent the height and width of the image, respectively; Step 1.3: Input the three-channel image E into the hierarchical feature fusion regression module. The hierarchical feature fusion regression module includes a trapezoidal CNN network. The trapezoidal CNN network has a total of N layers. Each layer has a corresponding number of bottleneck blocks. A VisionTransformer module is added after each bottleneck block in the last layer of the CNN network. After the three-channel image E enters the CNN network, the feature map F extracted from the i-th layer is obtained. i Feature alignment is performed using bottleneck blocks to obtain aligned feature maps. Alignment feature map With the feature map F of the next layer i+1 Then, the bottleneck block is used for fusion and alignment. The alignment features of the first N-1 layers are processed by the Vision Transformer module to obtain the features. The feature map F extracted from the last layer of the CNN network N With features The feature F is obtained by fusion, and the feature F is passed through a pooling layer to obtain the prediction quality score; Step 2: Construct the loss function L total Optimize the image quality assessment network, loss function L total Including MSE loss L quality Image relative ranking loss L relative .

2. The CT image quality assessment method based on the Poisson flow generation model according to claim 1, characterized in that, The difference diagram I e and dissimilar graph I dis The calculation method is as follows: I e =|I d -G θ (I d )| I dis =(1-SSIM(I d ,G θ (I d ))) In the above formula, θ represents the network parameters.

3. The CT image quality assessment method based on the Poisson flow generation model according to claim 1, characterized in that, The bottleneck block consists of a 1×1 convolutional layer, a 3×3 convolutional layer, and a 1×1 convolutional layer, aligned with the feature map. The calculation formula is as follows: Among them, W 1×1 and W 3×3 These represent the weight matrices for 1×1 and 3×3 convolutional layers, respectively. The number of bottleneck blocks in the i-th layer of the CNN network is i; The formula for calculating feature F is as follows:

4. The CT image quality assessment method based on the Poisson flow generation model according to claim 1, characterized in that, Loss function L total The calculation formula is as follows: L total =L quality +L relative The MSE loss is calculated as follows: In the above formula, B represents the total number of images in each batch, and s k and Let represent the true quality score and the predicted quality score of the k-th image, respectively; The relative ranking loss of the image is calculated as follows: L relative =L triplet (q max ,q′ max ,q min )+L triplet (q min ,q′ min ,q max ) =max{0,d(q′ max ,q max )-d(q max ,q min )+margin1}+max{0,d(q′ min ,q min )-d(q max ,q min )+margin2} Within each batch of CT images, q max ,q′ max q min ,q′ min These represent the predicted quality scores of the images with the highest, second-highest, lowest, and second-lowest subjective quality scores, respectively. Let q represent the predicted quality score. max The subjective quality score corresponding to the image is set. as well as

Images