Radiographic image quality evaluation and enhancement method and system based on deep learning

Abstract

The invention relates to a radiographic image quality evaluation and enhancement method and system based on deep learning. The method comprises the following steps: constructing a multi-modal radiographic image quality evaluation data set containing noise types, artifact levels and diagnostic availability labels; designing a double-branch deep neural network architecture containing a quality evaluation branch and an image enhancement branch, the quality evaluation branch fusing global and local features to evaluate image quality, and the image enhancement branch realizing local enhancement of a focus area; coordinating a loss function optimization process by adopting a self-adaptive joint training strategy; and deploying an online correction module to iteratively optimize the enhancement parameters. The system comprises an image acquisition interface module, a preprocessing module, a quality evaluation engine module, an intelligent enhancement module, a post-processing module and the like. The technology can accurately evaluate the quality of the radiation image, effectively enhance the image, reduce noise and artifact interference and improve the diagnosis availability, and has important application value in the field of medical image processing.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and specifically to a method and system for radiological image quality assessment and enhancement based on deep learning. Background Art

[0002] Radiological imaging plays a key role in modern medical diagnosis. Multimodal imaging technologies such as CT, MRI, and PET-CT provide doctors with a wealth of diagnostic information. However, the quality of radiological images in actual applications varies greatly, which seriously affects the accuracy and reliability of diagnosis.

[0003] Noise and artifact problems are prominent: images are easily disturbed by various noises, such as Gaussian noise, Poisson noise, etc. At the same time, motion artifacts and metal artifacts often appear. These noises and artifacts will blur the image details, interfere with the doctor's observation and judgment of the lesion, and increase the risk of misdiagnosis or missed diagnosis.

[0004] Traditional evaluation methods have limitations: In the past, image quality evaluation mostly relied on manual work, which was not only inefficient but also greatly influenced by subjective factors, and different doctors had different evaluation criteria. Traditional algorithms were also unable to cope with complex image features, making it difficult to accurately quantify image quality and unable to meet clinical needs for high-quality image evaluation.

[0005] Poor image enhancement: Existing image enhancement technologies have limited effects in improving image clarity and diagnostic usability. Some methods lose useful information when removing noise, or cause anatomical structure deformation during the enhancement process, and cannot balance noise suppression and detail preservation well, and cannot provide strong support for clinical diagnosis.

[0006] Lack of integrated solutions: In clinical practice, image evaluation and enhancement are often independent of each other, and there is a lack of a unified and efficient technical solution to organically combine the two. It is difficult to optimize the enhancement process in real time based on the image quality assessment results, and it is impossible to form a complete image quality improvement system, which restricts the application value of radiological images in clinical diagnosis.

[0007] Therefore, a radiological image quality assessment and enhancement method and system based on deep learning is needed to solve the above problems. Summary of the invention

[0008] In order to solve the problems of the prior art, the present invention provides a radiological image quality assessment and enhancement method and system based on deep learning.

[0009] In order to solve the above technical problems, the present invention is implemented by the following technical solutions: In the first aspect, a radiological image quality assessment and enhancement method based on deep learning comprises the following steps:

[0010] S1. Construct a multimodal radiological image quality assessment dataset, which contains medical image data annotated with noise type, artifact level and diagnostic usability labels;

[0011] Collect CT (slice thickness 0.5-5mm), MRI (1.5T / 3.0T), PET-CT multi-source imaging data, covering 8 major anatomical regions of the body

[0012] The labeling system includes:

[0013] Noise type: Gaussian noise (σ=0.05-0.2), Poisson noise (λ=5-50)

[0014] Artifact grade: three-level classification for motion artifacts, four-level severity score for metal artifacts Diagnostic usability Labeling: Labeled by three radiologists with a probability value of 0-1.

[0015] S2. Design a dual-branch deep neural network architecture, including:

[0016] Quality assessment branch: uses a multi-scale attention mechanism to fuse global features and local detail features, and outputs image quality scores and defect type classification results;

[0017] Quality Assessment Branch:

[0018] Use 3D Vision Transformer to extract global features (patch size 16×16×16)

[0019] Densely connected CNNs with 7×7×7 convolutional kernels deployed in parallel capture local details

[0020] Fusion of features through channel attention mechanism:

[0021]

[0022] Output quality score Q∈[0,1] and artifact type encoding.

[0023] Image enhancement branch: construct a residual dense connection network and achieve local enhancement of the lesion area through a deformable convolutional layer;

[0024] Construct a 10-layer residual dense network (growth rate = 32, 3×3×3 convolutions per layer)

[0025] Integrated deformable convolution layer, the deformation parameter δ is generated by interpolating the features of the adjacent five layers

[0026] The SSIM of the output enhanced image is ≥ 0.93.

[0027] S3. Design an adaptive joint training strategy to coordinate the optimization process of quality assessment loss function and image enhancement loss function through a dynamic weight allocation module;

[0028] Dynamic loss weight allocation formula:

[0029]

[0030] Training phase division:

[0031] Initial stage (t<0.3T): α∈[0.7,0.9] focuses on quality assessment

[0032] Mid-term (0.3T≤t<0.7T): Balance optimization

[0033] Late stage (t≥0.7T): β∈[0.7,0.9] focuses on the enhancement effect.

[0034] S4. Deploy an online correction module to iteratively optimize the enhancement parameters based on the quality assessment results of the enhanced image through a real-time feedback mechanism;

[0035] Create a parameter mapping matrix:

[0036]

[0037] Iteration termination condition:

[0038] Quality score Q ≥ 0.85 or

[0039] Iterations ≥ 5

[0040] A single iteration takes ≤83ms.

[0041] In this application, starting from building a data set, through network architecture design, training strategy formulation to online correction module deployment, a complete system is formed. Each step complements each other and plays a significant role in improving image quality and assisting diagnosis.

[0042] Construction of a multimodal radiological image quality assessment dataset: CT, MRI, and PET-CT multi-source image data are collected and labeled with noise types, artifact levels, and diagnostic usability labels, covering eight anatomical regions of the body. This provides a rich and standardized data foundation for subsequent research, ensuring that images of different modalities can be effectively analyzed, enabling the model to learn comprehensive and targeted image features, and facilitating accurate assessment of image quality and enhanced processing.

[0043] Dual-branch deep neural network architecture design

[0044] Quality assessment branch: Fusion of global and local features, output of image quality score and defect type classification results. Global and local features are extracted through 3D Vision Transformer and convolutional neural network respectively, and then fused through channel attention mechanism, or cross attention mechanism is used to achieve multi-modal feature fusion, output of more comprehensive assessment results, accurate location of image problems, and provide a basis for subsequent enhancement.

[0045] Image enhancement branch: Construct a residual densely connected network combined with a deformable convolutional layer, or adopt a multi-stage progressive enhancement architecture to achieve local enhancement of the lesion area, improve image clarity and diagnostic usability, and reduce the interference of noise and artifacts on diagnosis.

[0046] Adaptive joint training strategy: The dynamic weight allocation module coordinates the optimization process of quality assessment and image enhancement loss functions, focusing on different goals in stages. The initial stage of training focuses on quality assessment accuracy to help the model better learn the image quality assessment standard; the mid-term balance optimization ensures the coordinated development of the two tasks; the later stage focuses on the enhancement effect, so that the model can more effectively improve the image enhancement quality and improve the overall performance of the model on the basis of mastering the evaluation ability.

[0047] Online correction module deployment: Iteratively optimize the enhancement parameters according to the quality assessment results of the enhanced image, establish the parameter mapping matrix and set the iteration termination conditions. It can provide real-time feedback and adjust the enhancement process to ensure that the final output image quality reaches a high standard and meets clinical diagnosis needs, while improving the practicality and reliability of the algorithm.

[0048] In one implementation of the first aspect, the quality assessment branch in step S2 includes:

[0049] Use 3D Vision Transformer to extract global context features of images;

[0050] The convolutional neural network branches deployed in parallel extract local texture features;

[0051] Multimodal feature fusion is achieved through cross-attention mechanism;

[0052] The output includes three-dimensional evaluation results of noise level score (0-1), artifact severity grade (grade I-IV), and diagnostic availability probability.

[0053] By using 3D Vision Transformer to extract global context features of the image, and taking advantage of its advantage in capturing long-distance dependencies, the overall feature information of the image is obtained. Convolutional neural network branches are deployed in parallel to extract local texture features, and the convolution layer can effectively focus on local details of the image. Multimodal feature fusion is achieved through the cross-attention mechanism, and global and local features are deeply integrated to explore the association between different features. The final output includes three-dimensional evaluation results of noise level score (0-1), artifact severity grading (grades I-IV), and diagnostic availability probability, quantifying image quality from multiple dimensions.

[0054] The multimodal feature fusion method makes the evaluation results more comprehensive and accurate. It can accurately locate the quality problems such as noise and artifacts in the image and their severity. It can also give the probability of the image's usability for diagnosis, providing a detailed and reliable basis for subsequent image enhancement and diagnostic decisions.

[0055] In one implementation of the first aspect, the image enhancement branch adopts:

[0056] Multi-stage progressive enhancement architecture, each stage contains dense residual blocks and channel attention modules;

[0057] Dynamic feature selection mechanism, adaptively adjusting the activation strength of each enhancement stage according to the quality assessment results;

[0058] The anatomical structure is kept constrained, and the enhancement process is guided by the organ masks generated by the segmentation network.

[0059] By adopting a multi-stage progressive enhancement architecture, each stage contains dense residual blocks and channel attention modules. Dense residual blocks can promote the flow and reuse of features and improve the network's learning ability; the channel attention module assigns weights according to the importance of features and focuses on key information. Using a dynamic feature selection mechanism, the activation intensity of each enhancement stage is adaptively adjusted according to the quality assessment results, and the areas with prominent quality problems are enhanced. The anatomical structure retention constraint is introduced, and the organ mask generated by the segmentation network is used to guide the enhancement process to ensure that the organ anatomical structure is not destroyed when enhancing the image.

[0060] Multi-stage progressive enhancement combined with dynamic feature selection can gradually optimize different imaging problems at different stages and improve the targeted enhancement effect; anatomical structure constraints ensure the anatomical accuracy of the enhanced image, reduce interference with diagnosis, improve image clarity and diagnostic usability, and better meet clinical diagnostic needs.

[0061] In one implementation of the first aspect, the adaptive joint training strategy includes:

[0062] Construct a composite loss function:

[0063] L total =α·L QA +β·L IE +γ·L reg

[0064] Where L QA is the cross entropy loss for quality assessment, L IE To enhance the structural similarity loss of images, L reg is the anatomical structure distortion constraint;

[0065] The dynamic weight adjustment module automatically adjusts the α, β, and γ parameters according to the training stage, focusing on quality assessment accuracy in the early stage and on optimization of enhancement effects in the later stage.

[0066] By constructing a composite loss function, including the cross entropy loss for quality assessment, the structural similarity loss for enhanced images, and the anatomical distortion constraint term. The cross entropy loss is used to measure the difference between the quality assessment result and the true label, the structural similarity loss focuses on the structural similarity between the enhanced image and the original clear image, and the anatomical distortion constraint term prevents the anatomical structure from deforming during the enhancement process. During the training process, the dynamic weight adjustment module automatically adjusts the parameters of each loss term according to the training stage. The initial training focuses on the accuracy of quality assessment, and the later training focuses on the optimization of the enhancement effect.

[0067] The composite loss function comprehensively considers multiple key factors of quality assessment and image enhancement. The dynamic weight adjustment enables the model to focus on different goals at different training stages, effectively balancing the two tasks of quality assessment and image enhancement, and improving the overall performance of the model, enabling it to accurately assess image quality and enhance images with high quality.

[0068] In one implementation of the first aspect, the online correction module implements:

[0069] Establish a mapping relationship matrix between quality assessment results and enhancement parameters;

[0070] Dynamically adjust convolution kernel parameters, feature channel weights and other enhanced network parameters through the reinforcement learning framework;

[0071] Set the maximum number of iterations threshold and automatically terminate the optimization process when the evaluation result reaches the preset standard.

[0072] By establishing a mapping relationship matrix between quality assessment results and enhancement parameters, the relationship between assessment results and enhancement parameters is clarified. Through the reinforcement learning framework, the enhanced network parameters such as convolution kernel parameters and feature channel weights are dynamically adjusted, and the reinforcement learning feedback mechanism is used to continuously optimize the enhanced network according to the assessment results. The maximum number of iterations is set. When the assessment result reaches the preset standard (such as quality score Q ≥ 0.85 or number of iterations ≥ 5 times), the optimization process is automatically terminated to avoid excessive iterations.

[0073] It realizes the real-time feedback adjustment of enhancement parameters according to the enhanced image quality assessment results, improves the adaptability and flexibility of the algorithm, ensures that the final output image quality reaches a high standard and meets clinical diagnosis needs, and at the same time enhances the practicality and reliability of the algorithm.

[0074] The second aspect is the radiological image quality assessment and enhancement system based on deep learning, including:

[0075] Image acquisition interface module: supports standardized input of medical image formats such as DICOM and NIFTI;

[0076] Preprocessing module: performs grayscale normalization, organ region segmentation and abnormal slice detection;

[0077] Quality assessment engine: deploying the dual-branch assessment model described in claim 2 and outputting a structured quality report;

[0078] Intelligent enhancement module: integrating the multi-stage enhancement network described in claim 3, providing diagnosis-oriented enhancement mode selection;

[0079] Post-processing module: implement adaptive adjustment of window width and window position and metadata reconstruction in accordance with DICOM standards.

[0080] In one embodiment of the second aspect, the quality assessment engine comprises:

[0081] Interpretable visualization component: Generates heat maps to mark the main quality defect areas;

[0082] Graded warning module: triggers automatic alarm when IV-level artifact is detected or the probability of diagnosis availability is <0.7;

[0083] Knowledge base update mechanism: Continuously optimize the evaluation model through the federated learning framework.

[0084] The quality assessment engine includes an explainable visualization component that generates a heat map to mark the main quality defect areas and intuitively displays the areas with quality problems in the image; a graded warning module that triggers an automatic alarm when a grade IV artifact is detected or the probability of diagnostic availability is <0.7, promptly reminding doctors to pay attention to images with serious quality problems; a knowledge base update mechanism that continuously optimizes the evaluation model through a federated learning framework, and continuously improves model performance by using multi-party data while protecting data privacy.

[0085] The explainable visualization component allows doctors to quickly understand image quality problems; the graded warning module ensures that doctors will not miss images with serious quality problems, reducing the risk of misdiagnosis; the knowledge base update mechanism enables the evaluation model to keep pace with the times, adapt to different scenarios and data changes, and improve the accuracy and reliability of the evaluation.

[0086] In one implementation of the second aspect, the intelligence enhancement module implements:

[0087] Equipment adaptive enhancement: Establish a mapping relationship library between CT, MRI, and PET equipment parameters and enhancement strategies;

[0088] Dose optimization mode: targeted enhancement of low-dose images guided by noise distribution estimation;

[0089] Multi-protocol output: Synchronously generate standard window level images and lesion focus enhanced images.

[0090] By implementing device adaptive enhancement, a mapping relationship library between CT, MRI, and PET device parameters and enhancement strategies is established, and appropriate enhancement strategies are selected according to the characteristics of different devices; in dose optimization mode, targeted enhancement of low-dose images is performed based on noise distribution estimation, effectively suppressing noise in low-dose images while retaining useful information; multi-protocol output simultaneously generates standard window position images and lesion-focused enhanced images to meet different diagnostic needs.

[0091] Effect: Device adaptive enhancement improves the adaptability of enhancement strategies, dose optimization mode enhances the diagnostic value of low-dose images, and multi-protocol output provides doctors with richer imaging information, helping doctors make more accurate diagnostic decisions.

[0092] In a third aspect, an electronic device includes: a processor and a memory;

[0093] The memory is used to store one or more program instructions;

[0094] A processor is used to run one or more program instructions to execute the steps of the deep learning radiological image quality assessment and enhancement method.

[0095] In a fourth aspect, a computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a radiological image quality assessment and enhancement method based on deep learning.

[0096] The beneficial effects of the present invention are:

[0097] 1. The present invention collects multi-source imaging data of CT, MRI, and PET-CT, covering eight anatomical regions of the whole body, and annotates noise type, artifact level, and diagnostic usability labels. This provides a rich and standardized data foundation for subsequent research, enabling the model to learn comprehensive and targeted imaging features, which helps to accurately evaluate image quality and enhance processing;

[0098] 2. The quality assessment branch integrates global and local features, extracts global and local features through 3D Vision Transformer and convolutional neural network respectively, and then integrates them through channel attention mechanism or cross attention mechanism to output three-dimensional assessment results including noise level score, artifact severity grading, and diagnostic availability probability. Accurately locate the problems in the image and provide detailed and reliable basis for subsequent image enhancement and diagnostic decision-making;

[0099] 3. The image enhancement branch constructs a residual dense connection network combined with a deformable convolutional layer, or adopts a multi-stage progressive enhancement architecture. It can achieve local enhancement of the lesion area, reduce the interference of noise and artifacts on diagnosis, and improve image clarity and diagnostic usability. Multi-stage progressive enhancement combined with dynamic feature selection can gradually optimize different image problems; the anatomical structure is kept constrained to ensure the accuracy of the anatomical structure of the enhanced image;

[0100] 4. The adaptive joint training strategy constructs a composite loss function, including the cross entropy loss for quality assessment, the structural similarity loss for enhanced images, and the anatomical distortion constraint. Through the dynamic weight adjustment module, the focus is placed on quality assessment accuracy in the early stages of training, and on optimization of enhancement effects in the later stages, effectively balancing the two tasks of quality assessment and image enhancement, and improving the overall performance of the model;

[0101] 5. The online correction module establishes a mapping relationship matrix between quality assessment results and enhancement parameters, dynamically adjusts the enhancement network parameters through the reinforcement learning framework, and sets the maximum iteration threshold. It can provide real-time feedback and adjust the enhancement process to ensure that the final output image quality reaches a high standard, meets clinical diagnosis needs, and improves the practicality and reliability of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] Figure 1 It is a schematic diagram of the system structure of the present invention.

[0103] Figure 2 It is a schematic flow chart of the method of the present invention.

[0104] Figure 3 It is a schematic diagram of the data processing flow of the present invention.

[0105] Figure 4 It is a logical flow diagram of the present invention.

[0106] Figure 5 It is a user interaction schematic diagram of the present invention. DETAILED DESCRIPTION

[0107] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0108] like Figures 1 to 4 The deep learning-based radiological image quality assessment and enhancement method and system shown.

[0109] Construction of multimodal radiological image quality assessment dataset:

[0110] Implementation steps:

[0111] Data collection:

[0112] Collection sources: CT (slice thickness 0.5-5mm), MRI (1.5T / 3.0T), and PET-CT multi-source imaging data.

[0113] Anatomical area: covers 8 major areas including head, chest, abdomen, pelvis, spine, limbs, heart, and breast. Each type of image contains at least 1,000 data sets.

[0114] Labeling system:

[0115] Noise type: Gaussian noise (σ=0.05-0.2), Poisson noise (λ=5-50), simulated by the algorithm and manually verified.

[0116] Artifact Level:

[0117] Motion artifact: three-level classification (mild, moderate, severe).

[0118] Metal artifacts: four-level scoring (level I no artifacts, level IV severe artifacts).

[0119] Diagnostic availability: The images were annotated independently by three radiologists and the average probability value (0-1) was taken.

[0120] Statistics table:

[0121]

[0122] Dual-branch deep neural network architecture design

[0123] Quality Assessment Branch:

[0124] Global feature extraction: 3D Vision Transformer (ViT), input size 256×256×256, block size 16×16×16, embedding dimension 768.

[0125] Local feature extraction: Densely connected CNN with 7×7×7 convolution kernels, 5 layers in total, and 64 output channels in each layer.

[0126] Feature fusion: Cross-attention mechanism, calculates the spatial correlation between global and local features, and outputs a fused feature vector.

[0127] Output: 3D assessment results (noise score, artifact level, probability of diagnostic availability).

[0128] Image enhancement branch:

[0129] Network structure: 10-layer residual dense network, growth rate 32, 3×3×3 convolution per layer, integrated deformable convolution layer (the offset is generated by interpolation of features from the adjacent 5 layers).

[0130] Enhancement effect: The SSIM of the output enhanced image is ≥ 0.93 (compared with the clear reference image).

[0131] Performance index table:

[0132]

[0133] Adaptive joint training strategy

[0134] Dynamic weight allocation formula:

[0135] Training phase division:

[0136] Initial stage (t<0.3T): quality assessment loss weight α=0.9, enhancement loss weight β=0.1.

[0137] Middle period (0.3T≤t<0.7T): α=0.6, β=0.4.

[0138] Late stage (t≥0.7T): α=0.3, β=0.7.

[0139] Online correction module deployment

[0140] Implementation process:

[0141] Initialization parameters: Generate initial enhancement parameters (such as convolution kernel weights, channel attention coefficients) based on the quality assessment results.

[0142] Iterative Optimization:

[0143] Each iteration takes ≤83ms, and the maximum number of iterations is 5.

[0144] Termination condition: quality score Q ≥ 0.85 or the maximum number of iterations is reached.

[0145] Parameter mapping matrix: The enhanced parameters are updated through the reinforcement learning framework, and the mapping relationship matrix dimension is 256×256.

[0146] Optimization effect table:

[0147] Iterations Quality score Q (mean) SSIM (mean) Time consumption (ms / case) 1 0.72 0.89 83 3 0.81 0.92 249 5 0.87 0.94 415

[0148] System module implementation

[0149] Quality Assessment Engine:

[0150] Interpretable visualization: Generate heatmaps, annotate noise / artifact areas (such as Figure 1 ).

[0151] Graded warning: A red alarm is triggered when grade IV artifacts or diagnostic availability < 0.7.

[0152] Intelligent enhancement module:

[0153] Equipment adaptation: Pre-store CT / MRI / PET equipment parameter library to match the best enhancement strategy.

[0154] Multi-protocol output: Standard window level images (WW / WL=350 / 40) and lesion enhanced images are generated synchronously.

[0155] Implementation effect verification

[0156] Test data set: 200 clinical images (1 / 3 each of CT / MRI / PET-CT).

[0157] Results Statistics:

[0158]

[0159] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A radiological image quality assessment and enhancement method based on deep learning, characterized in that: The following steps are involved: S1. Construct a multimodal radiological image quality assessment dataset, wherein the dataset contains medical image data annotated with noise type, artifact level and diagnostic usability labels; S2. Design a dual-branch deep neural network architecture, including: Quality assessment branch: uses a multi-scale attention mechanism to fuse global features and local detail features, and outputs image quality scores and defect type classification results; Image enhancement branch: construct a residual dense connection network and achieve local enhancement of the lesion area through a deformable convolutional layer; S3. Design an adaptive joint training strategy to coordinate the optimization process of quality assessment loss function and image enhancement loss function through a dynamic weight allocation module; S4. Deploy an online correction module to iteratively optimize the enhancement parameters based on the quality assessment results of the enhanced image through a real-time feedback mechanism.

2. The method and system for radiological image quality assessment and enhancement based on deep learning according to claim 1, characterized in that: The quality assessment branch in step S2 includes: Use 3D Vision Transformer to extract global context features of images; The convolutional neural network branches deployed in parallel extract local texture features; Multimodal feature fusion is achieved through cross-attention mechanism; The output includes three-dimensional evaluation results of noise level score (0-1), artifact severity grade (grade I-IV), and diagnostic availability probability.

3. The method and system for radiological image quality assessment and enhancement based on deep learning according to claim 1, characterized in that: The image enhancement branch adopts: Multi-stage progressive enhancement architecture, each stage contains dense residual blocks and channel attention modules; Dynamic feature selection mechanism, adaptively adjusting the activation strength of each enhancement stage according to the quality assessment results; The anatomical structure is kept constrained, and the enhancement process is guided by the organ masks generated by the segmentation network.

4. The method and system for radiological image quality assessment and enhancement based on deep learning according to claim 1, characterized in that: The adaptive joint training strategy includes: Construct a composite loss function: L total =α·L QA +β·L IE +γ·L reg Where L QA is the cross entropy loss for quality assessment, L IE To enhance the structural similarity loss of images, L reg is the anatomical structure distortion constraint; The dynamic weight adjustment module automatically adjusts the α, β, and γ parameters according to the training stage, focusing on quality assessment accuracy in the early stage and on optimization of enhancement effects in the later stage.

5. The method and system for radiological image quality assessment and enhancement based on deep learning according to claim 1, characterized in that: The online correction module realizes: Establish a mapping relationship matrix between quality assessment results and enhancement parameters; Dynamically adjust convolution kernel parameters, feature channel weights and other enhanced network parameters through the reinforcement learning framework; Set the maximum number of iterations threshold and automatically terminate the optimization process when the evaluation result reaches the preset standard.

6. Radiological image quality assessment and enhancement system based on deep learning, characterized by include: Image acquisition interface module: supports standardized input of medical image formats such as DICOM and NIFTI; Preprocessing module: performs grayscale normalization, organ region segmentation and abnormal slice detection; Quality assessment engine: deploying the dual-branch assessment model described in claim 2 and outputting a structured quality report; Intelligent enhancement module: integrating the multi-stage enhancement network described in claim 3, providing diagnosis-oriented enhancement mode selection; Post-processing module: implement adaptive adjustment of window width and window position and metadata reconstruction in accordance with DICOM standards.

7. The radiological image quality assessment and enhancement system based on deep learning according to claim 6, characterized in that: The quality assessment engine comprises: Interpretable visualization component: Generates heat maps to mark the main quality defect areas; Graded warning module: triggers automatic alarm when IV-level artifact is detected or the probability of diagnosis availability is <0.7; Knowledge base update mechanism: Continuously optimize the evaluation model through the federated learning framework.

8. The radiological image quality assessment and enhancement system based on deep learning according to claim 6, characterized in that: The intelligent enhancement module realizes: Equipment adaptive enhancement: Establish a mapping relationship library between CT, MRI, and PET equipment parameters and enhancement strategies; Dose optimization mode: targeted enhancement of low-dose images guided by noise distribution estimation; Multi-protocol output: Synchronously generate standard window level images and lesion focus enhanced images.

9. An electronic device, characterized in that: The device comprises: a processor and a memory; The memory is used to store one or more program instructions; The processor is used to run one or more program instructions to execute the steps of the deep learning radiological image quality assessment and enhancement method as described in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the radiological image quality assessment and enhancement method based on deep learning as described in any one of claims 1 to 5.

Images