CT image quality evaluation method based on Poisson flow generation model

Abstract

The invention discloses a CT image quality evaluation method based on a Poisson flow generation model, and the method comprises the steps: generating a pseudo reference image through a pseudo reference image generation module PPFM, so as to make up for the missing of a reference image, helping a subsequent score regression model to capture detailed features, fusing the upper features and lower features together layer by layer through a hierarchical feature fusion regression module, and carrying out the recognition of the upper features and the lower features. The method comprises the following steps of: extracting a plurality of layers of features of different scales so as to make full use of the features of different scales, transmitting the last layer of extracted features to a Vision Transform module in sequence, and fusing the obtained global features and local features together so as to predict the final image quality.

Description

Technical Field

[0001] The invention belongs to the field of medical image and computer vision technology, and in particular relates to a CT image quality assessment method based on a Poisson flow generation model. Background Art

[0002] Computed tomography (CT) is one of the most commonly used medical imaging techniques and plays an indispensable role in the doctor's diagnostic process. However, the high-dose radiation during CT scanning can cause certain damage to the patient's body and pose a great health risk. Therefore, imaging at a lower radiation dose has become an important technology to reduce the risk of radiation exposure. However, while low-dose CT radiation is reduced, image quality is also reduced. The image contains factors such as noise and artifacts, which may interfere with the doctor's accurate diagnosis and affect subsequent image processing tasks such as image segmentation, restoration, and enhancement. Therefore, advancing accurate image quality assessment (IQA) methods is crucial to optimizing the radiation dose of CT scans, ensuring patient safety, and the reliability of diagnostic results.

[0003] For a given image, the goal of IQA is to predict a score as close to the mean subjective score (MOS) of humans as possible using an image quality assessment algorithm. Depending on whether a reference image is required, IQA can be divided 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 image and the distorted image to calculate the quality score, while partial-reference image quality assessment uses information from the distorted image and partial reference image information to calculate the quality score. Due to their reliance on reference images to calculate image quality scores, full-reference and partial-reference image quality assessments show good performance. Therefore, peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) are widely used as standard evaluation metrics for image quality because they are computationally efficient and have a strong correlation with human visual perception of traditional distortion types such as blur, compression, and Gaussian noise. However, in real-world scenarios, reference images are often not available, limiting the applicability of these methods in practical situations. This problem is particularly applicable in the medical field, where obtaining high-quality reference images may have adverse effects on patients. Therefore, no-reference image quality assessment methods do not require reference images to calculate image scores, are more suitable for practical applications, and are worthy of further development.

[0004] Traditional No-reference Image Quality Assessment (NR-IQA) methods mainly rely on the extraction of manual features to evaluate image quality by simulating the perceptual characteristics of the human visual system. After preprocessing the image, various manual features are extracted from the image, including statistical features (such as mean, variance), texture features (such as Gabor filter), edge features (such as Sobel operator), structural features (such as local binary pattern), and frequency domain features (such as Fourier transform). These features are fused through different methods to extract the statistical distribution of the features. 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 representation ability of manual features is often unsatisfactory.

[0005] Recently, convolutional neural networks (CNNs) have attracted widespread attention in NR-IQA tasks due to their powerful capabilities. Hyper-IQA divides features into low-level and high-level categories, and then transforms the latter to reshape the influence of the former. SAHN develops an adaptive super network to aggregate local distortion features and global semantic features. Meta-IQA proposes a meta-learning framework for IQA to improve the generalization ability of unknown image distortions. There are also some methods that add attention mechanisms to CNN to enrich local fine-grained details and multi-scale global semantic information. MANIQA proposes a multi-dimensional attention network for no-reference image quality assessment. MUSIQ proposes a multi-scale image quality Transformer, which uses the Transformer architecture to solve the problem of images of different sizes and aspect ratios.

[0006] At the same time, many researchers have also explored the field of medical image quality assessment. Due to the different characteristics of medical imaging, there are many different methods for specific imaging modes. These include medical image types such as MRI, ultrasound imaging, and X-ray. There are also many studies in the field of CT image quality assessment. SSIQA proposes a reference-free 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 of unlabeled data to guide model training. KFKT-IQA uses natural images of similar tasks for transfer learning. There are also some methods that use task-based IQA models to detect virtual inserted objects with simple geometric shapes and calculate the quantitative quality of CT images.

[0007] However, the lack of reference information poses a challenge, making these methods always have bottlenecks compared to FR-IQA and may have an adverse impact on their overall performance. Therefore, some recent methods adopt reconstruction-based strategies to learn quality-aware features from raw images. 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 hallucinated references conditioned on distorted images. Similarly, Ma et al. proposed to generate the main content of distorted images. In another study, Ren et al. proposed a restorative adversarial network to reconstruct the input distorted images. Recently, D-BIQA proposed to use 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, but not for CT images, because the noise in CT images is generally quantum noise that obeys Poisson noise, that is, the random fluctuations of photons will cause granular noise in the image. Summary of the invention

[0008] Purpose of the invention: The purpose of the present invention is to solve the deficiencies in the prior art and to provide 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 make up for the lack of a reference image and to help the subsequent score regression model capture detailed features. A hierarchical feature fusion regression module fuses the upper layer features with the lower layer features layer by layer to make full use of features of different scales. At the same time, the last layer of extracted features is sequentially passed to the ViT (Vision Transformer module), and the obtained global features are fused with the local features to predict the final image quality.

[0009] Technical solution: A CT image quality assessment method based on a Poisson flow generation model of the present invention comprises 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 contents are as follows:

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

[0012] Step 1.2: The original low-dose CT image I d and the corresponding difference map I e and dissimilar graph I dis Fusion 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, which includes a trapezoidal CNN network. The trapezoidal CNN network has a total of N layers, and a corresponding number of bottleneck blocks are set in each layer. A Vision Transformer module is added after each bottleneck block of the last layer of the CNN network.

[0014] After the three-channel image E enters the CNN network, the feature map F is extracted from the i-th layer i Perform feature alignment through the bottleneck block to obtain the aligned feature map Aligning feature maps and the feature map F of the next layer i+1 Then the bottleneck block is used for fusion alignment, and the alignment features of the first N-1 layers are processed by the Vision Transformer module. 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 the pooling layer to obtain the predicted quality score;

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

[0016] Further, the difference map 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 is the network parameter, SSIM() is the structural similarity measurement function, and SSIM() is used here to measure the similarity between two images.

[0020] Since the number of channels and dimensions of each layer of the CNN network are different, in order to integrate the features extracted from each layer into the final feature representation, a bottleneck block consisting of three convolution operations is set. The bottleneck block includes a 1×1 convolution layer, a 3×3 convolution layer, and a 1×1 convolution layer in sequence. The alignment feature The calculation formula is as follows:

[0021]

[0022] Among them, W 1×1 and W 3×3 Represent the weight matrices of the 1×1 and 3×3 convolutional layers respectively; so far, the feature map F i The number of channels and resolution are adjusted to appropriate levels;

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

[0024] The calculation formula of 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, s k and Represent the true quality score (subjective quality score) and predicted quality score of the kth 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] In each batch of CT images, q max ,q′ max ,q min ,q′ min Represent the predicted quality scores of the images with the highest, second highest, lowest, and second lowest subjective quality scores respectively;

[0035] Represents the predicted quality score as q max The subjective quality scores corresponding to the images of are given in Table 1. 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 a corresponding pseudo reference image based on the Poisson flow generation model. The smaller the gap between the pseudo reference image and the real reference image, the more accurate the judgment of the quality regression network is; this helps the subsequent score regression model to 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 a trapezoidal structure as a whole, hierarchically integrating features extracted from the middle layer, so that the model can fully utilize visual information features from low to high layers;

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

[0041] (4) The present invention introduces relative ranking loss to construct the overall loss function to fully utilize the relative relationship between data. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0044] Figure 3 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 It is an imaging comparison diagram of each technical solution in the embodiment. DETAILED DESCRIPTION

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

[0047] The CT image quality assessment network of the present invention includes two 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. According to research, two main factors currently affect the quality of CT images: noise and artifacts, among which quantum noise follows a Poisson distribution. To this end, the present invention uses a Poisson flow generation model to generate corresponding pseudo reference images to make up for the lack of reference images and help the subsequent score regression model capture detailed features. Experiments show that by adding pseudo reference images, the present invention can improve the performance of the model in all evaluation indicators. The hierarchical feature fusion regression module of the present invention adopts a ladder structure to fuse the upper-level features with the lower-level features layer by layer, and can make full use of features of different scales. At the same time, the obtained global features are fused with the local features through ViT 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 comprises 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 contents are as follows:

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

[0051] Step 1.2: The original low-dose CT image I d and the corresponding difference map I e and dissimilar graph I dis Fusion 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, which includes a trapezoidal CNN network. The trapezoidal CNN network has a total of N layers, and a corresponding number of bottleneck blocks are set in each layer. A Vision Transformer module is added after each bottleneck block of the last layer of the CNN network.

[0053] After the three-channel image E enters the CNN network, the feature map F is extracted from the i-th layer i Perform feature alignment through the bottleneck block to obtain the aligned feature map Aligning feature maps and the feature map F of the next layer i+1 Then the bottleneck block is used for fusion alignment, and the alignment features of the first N-1 layers are processed by the Vision Transformer module. 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 the pooling layer to obtain the predicted quality score;

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

[0055] Further, the difference map 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 is the network parameter.

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

[0060]

[0061] Among them, W 1×1 and W3×3 Represent the weight matrices of 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 calculation formula of feature F is as follows:

[0064] In order to further improve the feature modeling capability of the image quality assessment network, this embodiment introduces a corresponding number of Vision Transformer modules in the image quality assessment network. The Vision Transformer modules include components such as position encoding, attention mechanism, and multi-layer perceptron. The calculation method of the attention mechanism is as follows:

[0065]

[0066] The projection of query (Q), key (K) and value (V) is 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 by PPFM (such as Figure 2 The processed multi-channel image is input into the hierarchical feature fusion regression network (as shown in Figure 3 First, the multi-channel image is input into the preprocessing backbone network Resnet50 network, and the features extracted from each layer are aligned through the bottleneck network, and then the features of different layers are fused layer by layer. After five layers of feature extraction, the last layer of features is input into VIT to fuse the extracted local features with the global features; the features finally obtained are passed through the pooling layer to obtain the quality score predicted by the image quality assessment network.

[0068] In order to utilize the features extracted by the convolutional network, this embodiment uses a fully connected layer as a fusion layer to map the above 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, s k and Represent the true quality score (subjective quality score) and predicted quality score of the kth image respectively;

[0073] Although MSE loss is effective for quality prediction tasks, it does not explicitly consider the ranking and correlation between images. Since the goal of the present 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] In each batch of CT images, q max ,q′ max ,q min ,q′ min Represent the predicted quality scores of the images with the highest, second highest, lowest, and second lowest subjective quality scores respectively;

[0078] Represents the predicted quality score as q max The subjective quality scores corresponding to the images of are similar to those of the other images. as well as

[0079] Example 1

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

[0081] The dataset used for training is the Mayo low-dose CT dataset, a public clinical dataset containing CT images of 10 patients and does not contain the human perception score for each image. The data is divided into a training set containing the first 8 patients (a total of 4,800 slices) and a validation set containing the last 2 patients (a total of 1,136 slices). After completing the training process, the trained PPFM is used to obtain the corresponding pseudo-reference image from the LDCTBIQA2023 challenge dataset. The LDCTBIQA2023 challenge dataset includes a total of 1,000 distorted abdominal images that contain both noise and streak artifacts, generated at four different dose levels: 100%, 50%, 25%, and 10%. Of these images, 900 were assigned to the training phase, and the remaining 100 were reserved for the testing phase. The final human perception score for each image is determined by averaging the individual scores of five skilled radiologists. Subsequently, the corresponding difference maps and dissimilarity maps are calculated and used according to Figure 3 The multi-channel image is obtained by connecting them in a way.

[0082] The obtained multi-channel image is used as the input of the hierarchical feature fusion regression network. First, it is input into the pre-processed Resnet50 network to obtain the features F of different layers. i , N in this embodiment is set to 5, that is, the CNN network of this embodiment has five layers; then through the bottleneck network, 1×1, 3×3 and 1×1 convolution layers are used in sequence, and the feature map F i The number of channels and resolution are adjusted to Then The features of the next layer are sequentially fused to fully learn the features of different layers; finally, the features of the fifth layer are input into ViT to obtain the global features, which are then passed through the pooling layer to obtain the scores predicted by the present 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 cycles 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 example 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 the sum of the three (OverAll) as indicators for evaluating model performance. The comparison methods include CNN-based methods (DBCNN, HyperIQA, and QPT), Transformers-based methods (TReS, AHIQ, MANIQA, and MD-IQA) and D-BIQA, a method based on denoising diffusion probability model.

[0085] Table 1 Comparison 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 AHQ 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 The present invention 0.9815 0.9837 0.9113 2.8765

[0087] As shown in Table 1, the technical solution of the present invention outperforms the most advanced methods in SROCC, PLCC and overall evaluation on the LDCTIQA2023 database. Compared with the three CNN-based methods, the image quality assessment network of the present invention improves the overall score by at least 5%. Compared with other methods that add Transformers, the total 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 the present invention is higher than D-BIQA in all three indicators. In addition, the pseudo-reference image generation speed of PPFM in the technical solution of the present invention is nearly 490 times faster than that of the general diffusion model, as shown in Table 2.

[0088] Table 2 Ablation study of generative models

[0089] Generative Models Overall AIT Diffusion model 2.7767 94.531 PPFM of the present invention 2.8765 0.192

[0090] Example 3

[0091] This embodiment uses the technical solution of the present invention to visualize the CT images and heat maps 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, heat maps, MOS (GT) and scores predicted by the present invention. Figure 4 The heat map clearly highlights some contents, which are important parts that strongly affect human perception. When humans view images, the features of interest greatly affect human perception, so these features have a greater impact on the image quality assessment network of the present invention.

[0092] In order to evaluate the efficiency of each module in the network of the present invention, the present invention conducts an ablation experiment. The pre-trained ResNet50 is used as the backbone model of the image quality assessment network, and the effect of each individual component is analyzed by comparing SROCC, KROCC and OverAll.

[0093] Table 3 Ablation studies of various components

[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 sampled Poisson flow generation model (PPFM), and processes the pseudo reference image generated by PPFM as the network input. It can be seen that the SROCC and KROCC indicators are improved by 0.6% and 1.3%, respectively. Then, the effectiveness of the proposed hierarchical feature fusion regression network (HFFRN) is checked by connecting it with the output features of ResNet50. SROCC and KROCC are significantly improved by about 0.4% and 1.4%. When PPFM and HFFRN are added at the same time, the performance is further improved. Finally, the relative ranking loss (Lr) is added on this basis, and SROCC and KROCC are further improved to the highest values ​​of 98.37% and 91.13%, reaching 2.8765 overall.

Claims

1. A CT image quality assessment method based on a Poisson flow generation model, characterized in that: The following steps are involved: 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 contents are as follows: Step 1.1: The pseudo reference image generation module PPFM is based on the Poisson flow generation model and converts the original low-dose CT image I d Input the pseudo reference image generation module PPFM, and after denoising, the pseudo reference image G is obtained. θ (I d ); Step 1.2: The original low-dose CT image I d and the corresponding difference map I e and dissimilar graph I dis Fusion is performed 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, which includes a trapezoidal CNN network. The trapezoidal CNN network has a total of N layers, and a corresponding number of bottleneck blocks are set in each layer. 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 is extracted from the i-th layer i Perform feature alignment through the bottleneck block to obtain the aligned feature map Aligning feature maps and the feature map F of the next layer i+1 Then the bottleneck block is used for fusion alignment, and the alignment features of the first N-1 layers are processed by the Vision Transformer module. 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 the pooling layer to obtain the predicted 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 and image relative ranking loss L relative .

2. The CT image quality assessment method based on the Poisson flow generation model according to claim 1 is characterized in that: The difference map 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 is the network parameter.

3. The CT image quality assessment method based on the Poisson flow generation model according to claim 1, characterized in that: The bottleneck block includes a 1×1 convolutional layer, a 3×3 convolutional layer, and a 1×1 convolutional layer in sequence, and the alignment feature The calculation formula is as follows: Among them, W 1×1 and W 3×3 Represent the weight matrices of 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 calculation formula of 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, s k and Represent the true quality score and predicted quality score of the kth 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} In each batch of CT images, q max , q′ max ,q min , q′ min Represent the predicted quality scores of the images with the highest, second highest, lowest, and second lowest subjective quality scores respectively; Represents the predicted quality score as q max The subjective quality score corresponding to the image is set as well as

Images