Deep learning based multi-modal medical image quality inspection system

Abstract

The application discloses a multi-modal medical image quality inspection system based on deep learning, belongs to the technical field of medical image processing quality inspection, and comprises a data acquisition and extraction module, an evaluation calculation module and a result display module. After the evaluation calculation module receives first modality image data, a first texture feature, second modality image data, a second texture feature, third modality image data and a third texture feature, first modality image quality scores, second modality image quality scores and third modality image quality scores are obtained. According to the image quality scores of single modalities and the vector conversion of double modality related features, double modality image quality scores are obtained. Finally, based on the double modality image quality scores and the repeatedly fused texture features extracted after fusion, multi-modal stable features are obtained. The application comprehensively considers multi-factors, the correlation between double modalities and the fusion stability of single modalities, so as to comprehensively evaluate the overall quality of multi-modal images.

Description

Technical Field

[0001] This invention relates to the field of medical image processing quality inspection technology, and in particular to a multimodal medical image quality inspection system based on deep learning. Background Technology

[0002] In the field of medical diagnostics, the application of multimodal medical imaging technology is becoming increasingly widespread. Different modalities of medical imaging can provide information about different aspects of the human body. Among them, CT is good at displaying the anatomical structure of the human body, MRI has a strong ability to resolve soft tissues, and PET can reflect the metabolic function of the human body.

[0003] Current technologies for medical image processing quality control focus only on single-modality indicators, resulting in an insufficiently comprehensive assessment of image quality. With the development of medical imaging, some existing technologies have systems that fuse and perform quality control on multiple modalities of medical images. However, even when considering multiple modal indicators, different clinical application scenarios have different requirements for sharpness, noise level, and contrast, but some existing technologies still use fixed weights, thus failing to adjust them according to specific circumstances.

[0004] Some existing technologies often fail to fully consider the correlation between the two modal images when evaluating the quality of dual-modal image fusion. The images of different modalities have differences in feature space. If these differences are not considered, the fused image will suffer from information loss and redundancy.

[0005] In addition, some existing technologies neglect the stability of the fusion results when performing final quality checks on multimodal images. The multimodal image fusion process is affected by a variety of factors. If stability is not considered, the quality check results will fluctuate significantly, which will affect the accuracy of diagnosis. Summary of the Invention

[0006] The technical problem to be solved by this invention is that the existing technology has the shortcomings of lacking comprehensive consideration, ignoring the correlation between multiple modalities, and not considering the stability of fusion. To this end, we propose a multimodal medical image quality inspection system based on deep learning.

[0007] The technical solution mainly consists of a deep learning-based multimodal medical image quality inspection system, including a data acquisition and extraction module, an evaluation and calculation module, and a result display module.

[0008] The input terminals of the data acquisition and extraction module are respectively connected to the output terminals of the computed tomography (CT) scanner, magnetic resonance imaging (MRI) scanner, and positron emission tomography (PET) scanner, facilitating the simultaneous acquisition of input data from each device that forms a multimodal image.

[0009] The data acquisition and extraction module acquires the first modal image data transmitted from the computed tomography (CT) scanner, the second modal image data transmitted from the magnetic resonance imaging (MRI) device, and the third modal image data transmitted from the positron emission tomography (PET) device.

[0010] Extract the first texture features from the first modal image data, the second texture features from the second modal image data, and the third texture features from the third modal image data;

[0011] After receiving the first modal image data, the first texture feature, the second modal image data, the second texture feature, the third modal image data, and the third texture feature, the evaluation calculation module performs the following operations:

[0012] After performing a quality assessment on the first modal image data, the second modal image data, and the third modal image data respectively by weight allocation, the quality scores of the first modal image, the second modal image, and the third modal image are obtained.

[0013] The first texture feature, the second texture feature, and the third texture feature are transformed and evaluated in vector form for any two-modal combination to obtain bimodal related features for any two-modal combination.

[0014] Based on the first modal image quality score, the second modal image quality score, the third modal image quality score, and bimodal correlation features, bimodal image fusion of any two modal combinations is performed to obtain the bimodal image quality score of any two modal combinations.

[0015] After repeatedly fusing the first modal image data, the second modal image data, and the third modal image data m times, the fused texture features are extracted, vector transformation and evaluation are performed, and multimodal stable features are obtained.

[0016] Based on the aforementioned bimodal correlation features and multimodal stability features, multimodal images are fused to obtain a multimodal quality inspection score.

[0017] The results display module is responsible for displaying the results of the multimodal quality inspection scoring.

[0018] Preferably, the first modal image data includes computed tomography (CT) scan clarity, CT scan noise level, and CT scan contrast transmitted from the CT scan device.

[0019] The second modal image data includes the magnetic resonance imaging sharpness, magnetic resonance imaging noise level, and magnetic resonance imaging contrast transmitted from the magnetic resonance imaging device;

[0020] The third modality image data includes the positron emission tomography (PET) image clarity, PET noise level, and PET contrast transmitted from the PET imager.

[0021] Preferably, the evaluation calculation module includes a single-modal evaluation unit, a dual-modal evaluation unit, and a multi-modal evaluation unit.

[0022] Preferably, the single-modal evaluation unit obtains the first modal image quality score by summing the weighted values ​​of the computed tomography (CT) scan sharpness, the CT scan noise level, and the CT scan contrast.

[0023] The second modality image quality score is obtained by summing the weighted values ​​of the magnetic resonance imaging sharpness, the magnetic resonance imaging noise level, and the magnetic resonance imaging contrast.

[0024] The third modality image quality score is obtained by summing the weighted values ​​of the positron emission tomography (PET) image clarity, the PET image noise level, and the PET image contrast.

[0025] In the weighting of the first modal image quality score, the second modal image quality score, and the third modal image quality score, the total weighting value is 1, and the weighting is assigned separately according to the emphasis of the magnetic resonance imaging (MRI) sharpness, the MRI noise level, and the MRI contrast, as well as the emphasis of the positron emission tomography (PET) sharpness, the PET noise level, and the PET contrast.

[0026] The single-modal evaluation unit calculates the weight allocation of the first-modal image data, second-modal image data, and third-modal image data regarding sharpness, noise level, and contrast transmitted by the computed tomography (CT) scanner, magnetic resonance imaging (MRI) scanner, and positron emission tomography (PET) scanner, respectively. This allows for flexible weight allocation of each device for sharpness, noise level, and contrast, enabling flexible adjustment of the emphasis.

[0027] Preferably, the arbitrary two-mode combination includes a first mode combination, a second mode combination, and a third mode combination;

[0028] First modal combination: the computed tomography (CT) scanner and the magnetic resonance imaging (MRI) device;

[0029] The second modal combination: the computed tomography (CT) scanner and the positron emission tomography (PET) scanner;

[0030] The third modal combination: the magnetic resonance imaging device and the positron emission tomography device;

[0031] The dual-modal evaluation unit obtains the dual-modal image quality score for any two modal combinations based on the first modal combination, the second modal combination, and the third modal combination:

[0032] Based on the bimodal correlation features of the first modality combination obtained from the first texture features and the second texture features, and the first modality image quality score and the second modality image quality score, the bimodal image quality score of the first modality combination is obtained.

[0033] Based on the bimodal correlation features of the second modality combination obtained from the first texture features and the third texture features, and the first modality image quality score and the third modality image quality score, the bimodal image quality score of the second modality combination is obtained.

[0034] Based on the bimodal correlation features of the third modality combination obtained from the second texture features and the third texture features, and the first modality image quality score and the second modality image quality score, the bimodal image quality score of the third modality combination is obtained.

[0035] The bimodal evaluation unit can achieve the fusion of the correlation between all bimodalities in a multimodal environment by obtaining the bimodal image quality score of any two-modal combination.

[0036] Preferably, the steps for converting and evaluating the dual-modal correlation features into vector form are as follows:

[0037] S1: Extract the first texture feature, the second texture feature, and the third texture feature based on any two modal combinations;

[0038] S2: After extraction, convert it into vector form, i.e., v i → (x1, x2, ... x n ) and v i+1 → (y1, y2, ... y n );

[0039] S3: Obtain the vector form v i → (x1, x2, ... x n ) and v i+1 →(y1, y2, ... y n The vector dot product and the vector modulus;

[0040] S4: Obtain the bimodal correlation features of any two-mode combination by dividing the vector dot product by the vector modulus.

[0041] Preferably, the multimodal evaluation unit obtains average quality features based on the average degree of the bimodal image quality scores of the first modality combination, the second modality combination, and the third modality combination.

[0042] The average quality score of the dual-modal image is multiplied by the multimodal stable features to obtain the stable average quality features. The multimodal stable features are obtained by repeatedly fusing the first modal image data, the second modal image data, and the third modal image data input from the computed tomography (CT) scanner, magnetic resonance imaging (MRI) scanner, and positron emission tomography (PET) scanner, based on repeated fusion experiments. This ensures the accuracy of the stable average quality features during quality inspection.

[0043] Based on the average degree of difference between the bimodal image quality scores of the first modal combination, the second modal combination, and the third modal combination and the average bimodal image quality score, quality discrete features are obtained.

[0044] The stable average quality feature is added to the discrete quality feature to obtain the multimodal quality inspection score.

[0045] Preferably, the steps for acquiring the multimodal stability features are as follows:

[0046] S1: The first modal image data, the second modal image data, and the third modal image data are repeatedly fused m times to obtain m fusion results f1, f2, ... f m ;

[0047] S2: For m fusion results f1, f2, ... f m The fused texture features are extracted to obtain the feature vector f1. → f2 → , ...f m → ;

[0048] S3: Obtain the feature vector f1 → f2 → , ...f m → The mean and variance of the eigenvectors;

[0049] S4: Utilize The multimodal stability features are obtained.

[0050] The technical effects and advantages of this invention are as follows:

[0051] In this invention, the data acquisition and extraction module comprehensively and objectively evaluates the quality of a single modality image by acquiring and extracting the first modality image data, second modality image data, and third modality image data from the computed tomography (CT) scanner, magnetic resonance imaging (MRI) scanner, and positron emission tomography (PET) scanner, respectively, as well as the weight allocation. Furthermore, the weight allocation can be adjusted according to different clinical application scenarios and needs.

[0052] In the bimodal evaluation unit, bimodal correlation features of any two modal combinations are introduced. The bimodal correlation features are calculated using the cosine similarity method, which can accurately measure the correlation between the two modal images. When evaluating the fusion quality of bimodal images, the fusion quality is adjusted by combining modal correlation, thereby improving the accuracy of the evaluation.

[0053] In addition, the multimodal evaluation unit introduces multimodal stability features, which evaluate the stability of the fusion results by calculating the variance through repeated fusion experiments. In the final quality inspection score, the average quality is adjusted in combination with the multimodal stability features, which can comprehensively evaluate the overall quality of multimodal images. This comprehensive evaluation method can more accurately reflect the actual quality level of multimodal images than existing technologies. Attached Figure Description

[0054] Figure 1 This is a flowchart of the method for the multimodal medical image quality inspection system. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings and preferred embodiments.

[0056] Reference Figure 1 As shown, the present invention provides a technical solution: a multimodal medical image quality inspection system based on deep learning, including a data acquisition and extraction module, an evaluation and calculation module, and a result display module;

[0057] The input terminals of the data acquisition and extraction module are respectively connected to the output terminals of the computed tomography (CT) scanner, magnetic resonance imaging (MRI) scanner, and positron emission tomography (PET) scanner.

[0058] The data acquisition and extraction module acquires the first modal image data transmitted from the computed tomography (CT) scanner, the second modal image data transmitted from the magnetic resonance imaging (MRI) scanner, and the third modal image data transmitted from the positron emission tomography (PET) scanner.

[0059] The first modality of image data includes computed tomography (CT) resolution, CT noise level, and CT contrast transmitted from the computed tomography (CT) equipment.

[0060] The second modality of image data includes the magnetic resonance imaging sharpness, magnetic resonance imaging noise level, and magnetic resonance imaging contrast transmitted from the magnetic resonance imaging equipment.

[0061] The third modality of image data includes positron emission tomography (PET) image clarity, PET noise level, and PET contrast transmitted from the PET equipment.

[0062] Extract the first texture features from the first modal image data, the second texture features from the second modal image data, and the third texture features from the third modal image data;

[0063] After receiving the first modal image data, the first texture feature, the second modal image data, the second texture feature, the third modal image data, and the third texture feature, the evaluation calculation module performs the following operations:

[0064] After weighting and evaluating the quality of the first modal image data, the second modal image data, and the third modal image data, the quality scores of the first modal image, the second modal image, and the third modal image are obtained.

[0065] The first texture feature, the second texture feature, and the third texture feature are transformed and evaluated in vector form for any two-modal combination to obtain bimodal related features for any two-modal combination.

[0066] Based on the first modality image quality score, the second modality image quality score, the third modality image quality score, and the bimodal correlation features, bimodal image fusion is performed for any two modal combinations to obtain the bimodal image quality score for any two modal combinations.

[0067] After repeatedly fusing the first modal image data, the second modal image data, and the third modal image data m times, the fused texture features are extracted, vector transformation is performed, and evaluation is conducted to obtain multimodal stable features.

[0068] Based on bimodal correlation features and multimodal stability features, multimodal images are fused to obtain multimodal quality inspection scores;

[0069] The evaluation calculation module includes a single-modal evaluation unit, a dual-modal evaluation unit, and a multi-modal evaluation unit;

[0070] The results display module is responsible for displaying the results of the multimodal quality inspection scoring.

[0071] In this embodiment, the close cooperation between the data acquisition and extraction module, the evaluation and calculation module, the result display module, and each unit brings significant benefits to the deep learning-based multimodal medical image quality inspection system. The single-modal evaluation unit in the evaluation and calculation module acquires multimodal image data with the help of the data acquisition and extraction module, and obtains single-modal feature data after feature extraction. The dual-modal evaluation unit combines the modality correlation coefficient to optimize the fusion quality evaluation and improve the system's ability to judge the image fusion effect. The multimodal evaluation unit considers the fusion stability and accurately calculates the multimodal quality inspection score.

[0072] Reference Figure 1 As shown in this implementation scheme: the single-modal evaluation unit obtains the first modal image quality score by summing the weighted values ​​of computed tomography (CT) scan sharpness, CT scan noise level, and CT scan contrast.

[0073] The second modality image quality score is obtained by summing the weighted values ​​of magnetic resonance imaging (MRI) sharpness, MRI noise level, and MRI contrast.

[0074] The third modality image quality score is obtained by summing the weighted values ​​of positron emission tomography (PET) sharpness, PET noise level, and PET contrast.

[0075] In the weighting of the first modality image quality score, the second modality image quality score, and the third modality image quality score, the total weighting value is 1. The weighting is then assigned separately based on the emphasis on the sharpness, noise level, and contrast of magnetic resonance imaging (MRI), the emphasis on the sharpness, noise level, and contrast of MRI, and the emphasis on the sharpness, noise level, and contrast of positron emission tomography (PET).

[0076] Let computed tomography (CT) be labeled a, magnetic resonance imaging (MRI) be labeled b, and positron emission tomography (PET) be labeled c. The specific calculation formula for single-modal image quality scoring is as follows:

[0077] ;

[0078] in:

[0079] DF i The single-modal image quality score is given for any one of the following devices: computed tomography, magnetic resonance imaging, and positron emission tomography, where i = a, b, c;

[0080] Q iThe resolution of any one of the following devices: computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET).

[0081] DF i The single-modal image quality score is given for any one of the following devices: computed tomography, magnetic resonance imaging, and positron emission tomography, where i = a, b, c;

[0082] Q i The resolution of any one of the following devices: computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET).

[0083] ;

[0084] in:

[0085] RF 2i Let 2i be the bimodal image quality score for any two-modal combination, and 2i = (a, b) / (a, c) / (c, b);

[0086] DF i and DF i+1 These are the single-modal image quality scores for any two-modal combination, i.e., i+1=i=a, b, c;

[0087] g is the modal correlation coefficient reflecting the bimodal correlation characteristics. The steps for vector form transformation and evaluation of the bimodal correlation characteristics are as follows:

[0088] S1: Extract the first texture feature, the second texture feature, and the third texture feature based on any two modal combinations;

[0089] S2: After extraction, convert it into vector form, i.e., v i → (x1, x2, ... x n ) and v i+1 → (y1, y2, ... y n );

[0090] S3: Obtain vector form v i → (x1, x2, ... x n ) and v i+1 → (y1, y2, ... y n The vector dot product and vector modulus are as follows:

[0091] ;

[0092] ;

[0093] S4: Obtain the bimodal correlation features of any two-mode combination by dividing the vector dot product by the vector modulus, as follows:

[0094] .

[0095] In this embodiment, it is worth noting that a modal correlation coefficient g is introduced. This coefficient, calculated using the cosine similarity method, accurately measures the similarity and correlation between two modalities of images. In multimodal medical image fusion, the information contained in different modalities is both complementary and correlated. The adjustment of fusion quality through the modal correlation coefficient g makes the evaluation of fusion quality more consistent with reality, thereby improving the effect and quality of bimodal image fusion. Furthermore, a geometric mean is employed in this embodiment. and arithmetic mean Combination method, geometric mean It places greater emphasis on the overall quality of the two single-modal images, while the arithmetic mean... This reflects their average quality, and combining the two can evaluate the quality of bimodal image fusion from different perspectives.

[0096] Reference Figure 1 As shown in this implementation scheme: the multimodal evaluation unit obtains the average quality feature based on the average degree of the bimodal image quality scores of the first modality combination, the second modality combination, and the third modality combination.

[0097] The average quality score of the dual-modal image is multiplied by the multimodal stable features to obtain the stable average quality features;

[0098] Based on the average degree of difference between the bimodal image quality scores of the first modal combination, the second modal combination, and the third modal combination and the average bimodal image quality score, the quality discrete features are obtained.

[0099] The stable average quality feature is added to the discrete quality feature to obtain the multimodal quality inspection score;

[0100] The specific calculation formula for multimodal quality inspection scores is as follows:

[0101] ;

[0102] ZF represents the multimodal quality inspection score, and n represents the modality combination quantity. The calculation result is the average quality of the bimodal image quality score output by any two-modal combination;

[0103] The steps for obtaining multimodal stable features are as follows:

[0104] S1: Perform m repeated fusions on the first modal image data, the second modal image data, and the third modal image data to obtain m fusion results f1, f2, ... f m ;

[0105] S2: For m fusion results f1, f2, ... f m After fusion, texture features are extracted to obtain feature vector f1. → f2 → , ...f m → ;

[0106] S3: Obtain the feature vector f1 → f2 → , ...f m → The mean and variance of the eigenvectors are as follows:

[0107] ;

[0108] ;

[0109] f avg → f is the average value of the eigenvectors. i → For the eigenvector f1 → f2 → ,......f m → Any one of the feature vectors in it;

[0110] S4: Utilize To obtain multimodal stable features;

[0111] w is the stability coefficient that reflects the stability characteristics of the multimodal system.

[0112] In this embodiment, it is worth noting that a stability coefficient w is introduced. This coefficient is obtained by calculating the variance through multiple repeated fusion experiments and is used to measure the stability of the multimodal image fusion result. During the multimodal image fusion process, due to the complexity of the data and the uncertainty of the algorithm, the fusion result will have certain fluctuations. The stability coefficient w is used to measure the average quality of the multimodal images. Adjustments were made to fully consider the impact of fusion stability on the final quality inspection results, making the results more reliable and reducing errors caused by fusion instability. Furthermore, the multimodal quality inspection score (ZF) integrates multiple dual-modal image quality scores (RF). 2i By calculating the average mass and standard deviation The overall quality of the multimodal imagery was comprehensively evaluated, and its average quality was... It reflects the average level of multiple sets of dual-modal fusion, while the standard deviation This reflects the degree of dispersion of each fusion result. This comprehensive evaluation method can grasp the quality status of multimodal images from both macro and micro levels, so that the final quality inspection score can more accurately reflect the actual quality of multimodal images.

[0113] It should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should also be within the scope of protection of this invention.

Claims

1. A deep learning-based multimodal medical image quality inspection system, characterized in that: It includes a data acquisition and extraction module, an evaluation and calculation module, and a results display module; The input terminals of the data acquisition and extraction module are respectively connected to the output terminals of the computed tomography (CT) scanner, magnetic resonance imaging (MRI) scanner, and positron emission tomography (PET) scanner. The data acquisition and extraction module acquires the first modal image data transmitted from the computed tomography (CT) scanner, the second modal image data transmitted from the magnetic resonance imaging (MRI) device, and the third modal image data transmitted from the positron emission tomography (PET) device. Extract the first texture features from the first modal image data, the second texture features from the second modal image data, and the third texture features from the third modal image data; After receiving the first modal image data, the first texture feature, the second modal image data, the second texture feature, the third modal image data, and the third texture feature, the evaluation calculation module performs the following operations: After performing a quality assessment on the first modal image data, the second modal image data, and the third modal image data with weighted allocation, the quality scores of the first modal image, the second modal image, and the third modal image are obtained. The first texture feature, the second texture feature, and the third texture feature are transformed and evaluated in vector form for any two-modal combination to obtain bimodal related features for any two-modal combination. Based on the first modal image quality score, the second modal image quality score, the third modal image quality score, and the bimodal correlation features, bimodal image fusion of any two modal combinations is performed to obtain the bimodal image quality score of any two modal combinations. After repeatedly fusing the first modal image data, the second modal image data, and the third modal image data m times, the fused texture features are extracted, vector transformation and evaluation are performed, and multimodal stable features are obtained. Based on the aforementioned bimodal correlation features and multimodal stability features, multimodal images are fused to obtain a multimodal quality inspection score. The results display module is responsible for displaying the results of the multimodal quality inspection scoring; The steps for obtaining the multimodal stable features are as follows: S1: The first modal image data, the second modal image data, and the third modal image data are repeatedly fused m times to obtain m fusion results; S2: Extract the fused texture features from the m fusion results to obtain feature vectors; S3: Obtain the mean and variance of the eigenvectors of the eigenvectors; S4: Utilize The multimodal stability features are obtained.

2. The deep learning-based multimodal medical image quality inspection system according to claim 1, characterized in that: The first modal image data includes computed tomography (CT) resolution, CT noise level, and CT contrast transmitted from the CT scanner. The second modal image data includes the magnetic resonance imaging sharpness, magnetic resonance imaging noise level, and magnetic resonance imaging contrast transmitted from the magnetic resonance imaging device; The third modality image data includes the positron emission tomography (PET) image clarity, PET noise level, and PET contrast transmitted from the PET imager.

3. The deep learning-based multimodal medical image quality inspection system according to claim 2, characterized in that: The evaluation calculation module includes a single-modal evaluation unit, a dual-modal evaluation unit, and a multi-modal evaluation unit.

4. The deep learning-based multimodal medical image quality inspection system according to claim 3, characterized in that: The single-modal evaluation unit obtains the first modal image quality score by summing the weighted values ​​of the computed tomography (CT) scan sharpness, the CT scan noise level, and the CT scan contrast. The second modality image quality score is obtained by summing the weighted values ​​of the magnetic resonance imaging sharpness, the magnetic resonance imaging noise level, and the magnetic resonance imaging contrast. The third modality image quality score is obtained by summing the weighted values ​​of the positron emission tomography (PET) image clarity, the PET image noise level, and the PET image contrast. In the weighting of the first modality image quality score, the second modality image quality score, and the third modality image quality score, the total weight allocation value is 1, and the weights are respectively allocated according to the emphasis of the magnetic resonance imaging (MRI) sharpness, the MRI noise level, and the MRI contrast, as well as the emphasis of the positron emission tomography (PET) sharpness, the PET noise level, and the PET contrast.

5. The deep learning-based multimodal medical image quality inspection system according to claim 4, characterized in that: The arbitrary two-mode combination includes a first mode combination, a second mode combination, and a third mode combination; First modal combination: the computed tomography (CT) scanner and the magnetic resonance imaging (MRI) device; The second modal combination: the computed tomography (CT) scanner and the positron emission tomography (PET) scanner; The third modal combination: the magnetic resonance imaging device and the positron emission tomography device; The dual-modal evaluation unit obtains the dual-modal image quality score for any two modal combinations based on the first modal combination, the second modal combination, and the third modal combination: Based on the bimodal correlation features of the first modality combination, and the first modality image quality score and the second modality image quality score, the bimodal image quality score of the first modality combination is obtained; Based on the bimodal correlation features of the second modality combination, and the first modality image quality score and the third modality image quality score, the bimodal image quality score of the second modality combination is obtained; The bimodal image quality score of the third modality combination is obtained based on the bimodal correlation features of the third modality combination, the first modality image quality score, and the second modality image quality score.

6. The deep learning-based multimodal medical image quality inspection system according to claim 5, characterized in that: The steps for vector form conversion and evaluation of the dual-modal correlation features are as follows: S1: Extract the first texture feature, the second texture feature, and the third texture feature based on any two modal combinations; S2: Extract and convert to vector form; S3: Obtain the vector dot product and vector modulus in the vector form; S4: Obtain the bimodal correlation features of any two-mode combination by dividing the vector dot product by the vector modulus.

7. The deep learning-based multimodal medical image quality inspection system according to claim 5, characterized in that: The multimodal evaluation unit obtains average quality features based on the average degree of the bimodal image quality scores of the first modality combination, the second modality combination, and the third modality combination. The average quality score of the dual-modal image is multiplied by the multimodal stable feature to obtain the stable average quality feature; Based on the average degree of difference between the bimodal image quality scores of the first modal combination, the second modal combination, and the third modal combination and the average bimodal image quality score, quality discrete features are obtained. The stable average quality feature is added to the discrete quality feature to obtain the multimodal quality inspection score.

Images