A Medical Image Causal Rationality Detection Method Based on Dual-Channel Condition Fusion

Abstract

The present invention discloses a method for detecting the causal rationality of medical images based on dual-channel conditional fusion, including: obtaining a medical image data set and selecting a medical image therefrom as the original medical image; using a medical image perturbator to obtain positive sample pairs and negative sample pairs; establishing a dual-channel conditional fusion model, and using the dual channels to respectively fuse two medical images in the positive sample pairs and negative sample pairs to extract their causal features. According to the time interval between the two medical images, the abstract features from the two medical images are compared and fused differently, and the differences caused by different patient postures and device parameters can be suppressed, so that the extracted abstract features are highly sensitive to time, the time perception ability is increased, the causal features can be better obtained, the problem that the time span in the prior art causal relationship extraction method is not adjustable and is too sensitive to the interference caused by different patient postures and imaging parameters is overcome, and more accurate detection results can be obtained.

Description

Technical Field

[0001] The present invention belongs to the field of medical image detection, and particularly relates to a method for detecting the causal rationality of medical images based on dual-channel conditional fusion. Background Art

[0002] With the remote, multi-modal fusion, and intelligent development of current medical imaging equipment, it is difficult to absolutely guarantee the quality and reliability of medical imaging data. The upload of medical images may be the upload of others', patients' past, algorithmically synthesized or tampered, or misoperation-caused incorrect imaging, either intentionally or unintentionally, thus affecting the doctor's diagnosis results. And with the intelligence of current medical imaging equipment, especially for some large-scale, wide-range, and screening medical images, due to their extremely large quantity, it is difficult to allocate professional personnel to review all uploaded images. Given that the tissues and organs in medical images have their fixed development laws over time, there must be a causal relationship between two medical images with time sequence. Specifically, biological tissues, organs, and even diseases also have their specific development processes and will not suddenly show changes that violate the normal laws of tissue and organ growth or disease development. Usually, artificial intelligence algorithms are used to review the uploaded images, and the temporal causality is used to judge whether the medical images belong to the target patient, whether they belong to historical data, and whether the changes in the tissues and organs reflected in the images conform to their natural development laws, so as to evaluate the causal rationality of tissues and organs in two images with time sequence without relying on manual review. However, when the patient undergoes two screenings successively, due to different patient postures and imaging parameters, some additional changes will be brought to the medical images, thus interfering with the extraction of the causal relationship.

[0003] Zhou Tao et al. proposed the Hi-net multi-modal hybrid medical image synthesis network structure. References: Zhou T, Fu H, Chen G, Shen J, Shao L. Hi-net: hybrid-fusion network for multi-modal MR image synthesis. IEEE transactions on medical imaging 2020, 39(9): 2772-2781. Hi-net uses two modalities of medical images to synthesize the third modality of medical images and sets up a hybrid fusion module to achieve information fusion at different abstraction levels from low to high. The two modalities of medical images as inputs can be two co-modal medical images with a causal relationship, and the causal features of the two medical images can be extracted through multi-modal fusion. However, it does not support the extraction of causal relationships under different time spans and is easily interfered by different patient postures and imaging parameters. Summary of the Invention

[0004] The purpose of the present invention is to address the above problems and propose a method for detecting the causal rationality of medical images based on dual-channel conditional fusion, which overcomes the problems in the prior art's causal relationship extraction methods, such as the non-adjustable time span and excessive sensitivity to interference caused by different patient postures and imaging parameters, and can obtain more accurate detection results.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] A method for detecting the causal rationality of medical images based on dual-channel conditional fusion proposed by the present invention includes the following steps:

[0007] S1. Obtain a medical image dataset and select a medical image from the medical image dataset as the original medical image where A is the patient number, i is the shooting and generation date of the original medical image, a represents the shooting part, and b represents the shooting angle;

[0008] S2. Use a medical image perturbator to select two medical images from the medical image dataset, denoted as the first medical image and the second medical image such that constitute a positive sample pair that conforms to causal rationality, constitute a negative sample pair that does not conform to causal rationality, Y is the time interval between the two medical images in the sample pair, B is the patient number, j is the shooting and generation date of the medical image selected by the medical image perturbator, c represents the shooting part, and d represents the shooting angle;

[0009] S3. Establish a dual-channel conditional fusion model. The dual-channel conditional fusion model includes a first channel module, a second channel module, a discriminator, and three feature extractors. Both the first channel module and the second channel module include a time information distributor and multiple sequentially connected embedding fusion units. The embedding fusion unit includes a time information embedder and a fusion module connected in sequence, and the following operations are performed:

[0010] S31. Input the original medical image the first medical image and the second medical image one by one into the three feature extractors for feature extraction to obtain the abstract features of each level with gradually increasing abstract levels corresponding thereto;

[0011] S32. Input the abstract features of each level with gradually increasing abstract levels corresponding to the original medical image and the first medical image into the respective time information embedders of the first channel module one by one in sequence, and input the abstract features of each level with gradually increasing abstract levels corresponding to the original medical image and the second medical image The abstract features at each level corresponding to the abstract levels from low to high are sequentially input into the respective time information embedders of the second channel module one by one;

[0012] S33. Use the time information distributor to convert the time interval Y between the two medical images of the sample pair into the mean and variance matching each level and distribute them to the time information embedders of the corresponding channel modules;

[0013] S34. Integrate the mean and variance received by the embedding fusion unit of each channel module into the abstract features matching it to obtain the output features of the corresponding channel module. The specific operations of each channel module are as follows:

[0014] S341. Integrate the abstract features of the two medical images received by the time information embedder at the first level with the received mean and variance respectively to obtain the corresponding first-level extraction features, and generate the first-level fusion features through the first-level fusion module for the first-level extraction features;

[0015] S342. Integrate the abstract features of the two medical images received by the time information embedder at the nth level with the received mean and variance respectively to obtain the corresponding nth-level extraction features, and generate the nth-level fusion features through the nth-level fusion module for the nth-level extraction features and the (n - 1)th-level fusion features, where n ≥ 2;

[0016] S343. Determine whether n is equal to N, where N is the number of embedding fusion units. If so, it is considered that the fusion of all levels is completed, and the fusion features at the last level are used as the output features of the corresponding channel module. Otherwise, set n = n + 1 and return to execute step S342;

[0017] S4. Record the output features of the first channel module and the output features of the second channel module as positive samples and negative samples in sequence and input them into the discriminator to obtain the causal rationality score of the original medical image of.

[0018] Preferably, the positive sample pairs conforming to causal rationality satisfy the following constraints:

[0019] Y is the time interval between the original medical image and the first medical image in years; j ≥ i, where both i and j are integers; when i = j, ab ≠ cd;

[0020] The negative sample pairs not conforming to causal rationality satisfy the following constraints:

[0021] ① When A = B: Y is the time interval between the original medical image and the second medical image The time interval, in years; i ≥ j, where both i and j are integers;

[0022] ② When A ≠ B: Y is the original medical image and the second medical image The time interval, in years; both i and j are integers.

[0023] Preferably, both the feature extractor and the discriminator adopt a multi-layer convolutional neural network.

[0024] Preferably, the feature extractor adopts a ResNet50 network or a Transformer network.

[0025] Preferably, the time information distributor adopts a fully connected network.

[0026] Preferably, the time information embedder is a conditional normalization layer.

[0027] Preferably, the conditional normalization layer adopts a Batch Normalization layer.

[0028] Preferably, the fusion method of the fusion module is one of multiplying, adding, and taking the maximum value of the input features in a pixel-level matrix.

[0029] Preferably, the dual-channel conditional fusion model is also trained and backpropagated to optimize the dual-channel conditional fusion model through the difference loss function Loss dif and the distribution loss function Loss dis until the difference loss function Loss dif and the distribution loss function Loss dis converges to obtain a trained dual-channel conditional fusion model.

[0030] Preferably, the difference loss function Loss dif and the distribution loss function Loss dis , the formula is as follows:

[0031]

[0032] Loss dis = -DD(Di(x))

[0033] In the formula, x is the positive sample, y is the negative sample, Di(·) is the difference discriminator, is the gradient of Di(x), ||·||1 is the L1 norm, is the square of the L2 norm, and DD(·) is the distribution discriminator.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] This method introduces a dual-channel conditional fusion model, which fuses two causally related images through dual channels and extracts their causal features. It can differentially compare and fuse the abstract features from two medical images according to the time interval between the two medical images, and can suppress the differences caused by different patient postures and device parameters, especially the low-level abstract features interfered by patient postures and imaging parameters. By adding a conditional normalization layer, the extracted abstract features are highly sensitive to time, increasing the network's time perception ability. And through the fusion module, as the training progresses, it can automatically learn to use the features of the required abstract level, better obtain causal features, overcome the problems in the prior art's causal relationship extraction methods, such as the non-adjustable time span and over-sensitivity to the interference caused by different patient postures and imaging parameters, and can obtain more accurate detection results. Description of the Drawings

[0036] Figure 1 It is a flowchart of the medical image causal rationality detection method based on dual-channel conditional fusion of the present invention. Detailed Embodiments

[0037] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0038] It should be noted that unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which this application belongs. The terms used in the specification of this application herein are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0039] To overcome the problems in the prior art's medical image causal relationship extraction methods, such as the non-adjustable time span and over-sensitivity to the interference caused by different patient postures and imaging parameters. The present invention provides a medical image causal rationality detection method based on dual-channel conditional fusion, which can differentially compare and fuse the features from two medical images according to the time interval between the two medical image acquisitions, and can suppress the differences caused by different patient postures and device parameters, and obtain more accurate detection results.

[0040] As Figure 1 shown, a medical image causal rationality detection method based on dual-channel conditional fusion includes the following steps:

[0041] S1. Obtain a medical image dataset and select a medical image from the medical image dataset and denote it as the original medical image Among them, A is the patient number, i is the shooting and generation date of the original medical image, a represents the shooting part, and b represents the shooting angle.

[0042] S2. Use a medical image perturbator to select two medical images from the medical image dataset and denote them as the first medical image and the second medical image such that constitute a positive sample pair that conforms to causal rationality, constitute a negative sample pair that does not conform to causal rationality, Y is the time interval between the two medical images in the sample pair, B is the patient number, j is the shooting and generation date of the medical image selected by the medical image perturbator, c represents the shooting part, and d represents the shooting angle.

[0043] In one embodiment, the positive sample pair that conforms to causal rationality satisfies the following constraints:

[0044] Y is the time interval between the original medical image and the first medical image , in years; j≥i, both i and j are integers; when i = j, ab≠cd;

[0045] The negative sample pair that does not conform to causal rationality satisfies the following constraints:

[0046] ① When A = B: Y is the time interval between the original medical image and the second medical image , in years; i≥j, both i and j are integers;

[0047] ② When A≠B: Y is the time interval between the original medical image and the second medical image , in years; both i and j are integers.

[0048] Among them, a and c represent different or the same shooting parts; b and d represent different or the same shooting angles. The medical image perturbator is used to judge whether the information of the two medical images traversed by for loop, etc. in the medical image dataset, such as patient number, shooting time, shooting part, shooting angle, etc., conforms to the definition of causal rationality according to the above constraints, and finally output. It can find the paired medical images that are reasonable or unreasonable in time and space for the given original medical image from the medical image dataset, and output the positive sample pairs that conform to causal rationality and the negative sample pairs that do not conform to causal rationality. It should be noted that the unit of the time interval can also be adjusted according to actual needs, such as months, days, etc.

[0049] S3. Establish a dual-channel conditional fusion model. The dual-channel conditional fusion model includes a first-channel module, a second-channel module, a discriminator, and three feature extractors. Both the first-channel module and the second-channel module include a time information distributor and multiple sequentially connected embedding fusion units. The embedding fusion unit includes a time information embedder and a fusion module connected in sequence, and performs the following operations:

[0050] S31. Input the original medical image the first medical image and the second medical image into the three feature extractors one by one for feature extraction to obtain abstract features at each level with corresponding abstract levels from low to high;

[0051] S32. Input the original medical image and the abstract features at each level with corresponding abstract levels from low to high of the first medical image into the respective time information embedders of the first-channel module one by one in sequence. Input the original medical image and the abstract features at each level with corresponding abstract levels from low to high of the second medical image into the respective time information embedders of the second-channel module one by one in sequence;

[0052] S33. Use the time information distributor to convert the time interval Y between the two medical images in the sample pair into a mean and variance matching each level and distribute them to the time information embedders of the corresponding channel modules;

[0053] S34. Integrate the mean and variance received by the embedding fusion units of each channel module into the abstract features matching them to obtain the output features of the corresponding channel modules. The specific operations of each channel module are as follows:

[0054] S341. Integrate the abstract features of the two medical images received by the time information embedder at the first level with the received mean and variance respectively to obtain the corresponding first-level extraction features, and generate first-level fusion features through the first-level fusion module with the first-level extraction features;

[0055] S342. Integrate the abstract features of the two medical images received by the time information embedder at the nth level with the received mean and variance respectively to obtain the corresponding nth-level extraction features, and generate nth-level fusion features through the nth-level fusion module with the nth-level extraction features and the (n - 1)th-level fusion features, where n ≥ 2;

[0056] S343. Determine whether n is equal to N, where N is the number of embedded fusion units. If so, it is considered that the fusion of all levels is completed, and the fusion feature of the last level is used as the output feature of the corresponding channel module. Otherwise, set n = n + 1, and return to execute step S342.

[0057] In one embodiment, both the feature extractor and the discriminator adopt multi-layer convolutional neural networks. It should be noted that the feature extractor is not limited to using convolutional neural networks to extract image features, and the methods for extracting features at each scale can also be adjusted according to actual needs.

[0058] In one embodiment, the feature extractor adopts a ResNet50 network or a Transformer network. Abstract features at various levels with gradually increasing abstraction levels are extracted from the input medical image, and the extracted abstract features at each level will be used as the input of the corresponding layer's time information embedder.

[0059] In one embodiment, the time information distributor adopts a fully connected network. The time interval Y is used as its input, and finally, the fully connected network predicts and outputs the mean and variance required by the time information embedder that matches each level.

[0060] In one embodiment, the time information embedder is a conditional normalization layer.

[0061] In one embodiment, the conditional normalization layer adopts a Batch Normalization layer.

[0062] It should be noted that other normalization layers in the prior art can also be adopted for the conditional normalization layer. After inputting the mean and variance input by the time information embedder into the conditional normalization layer, they are fused with the extracted abstract features of the corresponding level, so that they present statistical characteristics matching the time and the abstract level, enhancing the time sensitivity of the extracted abstract features and the time perception ability of the network.

[0063] In one embodiment, the fusion method of the fusion module is to perform one of matrix multiplication, addition, and taking the maximum value of pixel-level matrices on its input features. Except for the fusion module of the first level that fuses the extracted features of this level, the fusion modules of other levels fuse the extracted features of this level and the fusion features output by the previous level fusion module. The fusion method can be matrix multiplication, addition, taking the maximum value, etc. of pixel-level matrices. Finally, the output features of the corresponding channels effectively integrate multi-level representations, improving the fusion performance.

[0064] In one embodiment, the dual-channel conditional fusion model is also trained and passed through the difference loss function Loss dif and the distribution loss function Loss disBackpropagate to optimize the dual-channel conditional fusion model until the difference loss function Loss dif and the distribution loss function Loss dis converge to obtain a trained dual-channel conditional fusion model.

[0065] In one embodiment, the difference loss function Loss dif and the distribution loss function Loss dis , the formula is as follows:

[0066]

[0067] Loss dis = -DD(Di(x))

[0068] where x is the positive sample, y is the negative sample, Di(·) is the difference discriminator, is the gradient of Di(x), ||·||1 is the L1 norm, is the square of the L2 norm, and DD(·) is the distribution discriminator. The difference discriminator and the distribution discriminator can be discriminator structures in the prior art or designed according to actual needs, such as using the discriminator structure in the WGAN network.

[0069] Loss dif can make the discriminator have as large a scoring difference between positive and negative samples as possible, but can avoid polarization leading to non-convergence. Loss dis can make the score of the positive sample satisfy a predefined random distribution, ensuring a high degree of aggregation of the scores of the positive samples to control the score distribution of the positive samples.

[0070] As the entire dual-channel conditional fusion model is trained, the fusion module can automatically learn which features at which abstraction levels can better obtain causal features, thereby effectively suppressing the interference caused by low-abstraction-level features to poses, configurations, etc.

[0071] S4. Record the output features of the first channel module and the output features of the second channel module as positive and negative samples in sequence and input them into the discriminator to obtain the causal rationality score of the original medical image . Use the features finally extracted and fused by the positive sample as the positive sample of the discriminator, use the features finally extracted and fused by the negative sample as the negative sample of the discriminator, and the dual-channel conditional fusion model optimized through iterative training can finally output the causal rationality score of the medical image.

[0072] The following is a detailed description through specific embodiments.

[0073] As shown in Figure 1 , the label description is as follows: represents the original medical image, represents the first medical image, that is, the medical image selected by the medical image disturber from the medical image dataset and paired as a positive sample pair of medical images, represents the second medical image, that is, the medical image selected by the medical image disturber from the medical image dataset and paired as a negative sample pair of medical images, G represents the medical image disturber, Y represents the time interval between the two medical images in the sample pair, E represents the time information distributor, and F0 to F9 represent the abstract features extracted by the feature extractor with the abstract levels from low to high, F 10 ~F 19 represents the abstract features extracted by the feature extractor with the abstract levels from low to high, F 20 ~F 29 represents the abstract features extracted by the feature extractor with the abstract levels from low to high, T0 to T9 represent the time information embedders corresponding to the corresponding levels (from low to high) in the first channel module, T 10 ~T 19 represents the time information embedders corresponding to the corresponding levels (from low to high) in the second channel module, M0 to M9 represent the fusion modules corresponding to the corresponding levels (from low to high) in the first channel module, M 10 ~M 19 represents the fusion modules corresponding to the corresponding levels (from low to high) in the second channel module, Net represents the feature extractor, and D represents the discriminator. In the figure, the thin solid arrows represent the positive sample pair data flow, the thick solid arrows represent the negative sample pair data flow, the thin dotted arrows represent the fusion data flow, and the thin dashed arrows represent the time information flow.

[0074] 1), Pass the original medical image through the medical image disturber G to obtain a positive sample pair that conforms to causal rationality and

[0075] a negative sample pair that does not conform to causal rationality and 2), Use two identical feature extractors Net to respectively obtain the abstract features F0 to F9 and F with the abstract levels from low to high for the two images of the positive sample pair 10 ~F 19 , The parameters are shared between the two feature extractors Net. Each level of abstract feature will have a corresponding time information embedder. For example, the abstract features of the first level in the first channel module include F0 and F 10 , and the corresponding time information embedder is T0. The same applies to others. That is, F0 to F9 and F of the positive sample pair 10 ~F 19The corresponding time information embedders are T0 to T9, the F0 to F9 of the negative sample pairs, and F 20 ~F 29 The corresponding time information embedders are T 10 ~T 19 。In this embodiment, a total of 19 time information embedders T1, T2, …, T 19 are used, corresponding to 19 fusion modules M1, M2, …, M 19 It should be noted that the number of time information embedders and fusion modules can also be adjusted according to actual needs.

[0076] 3), The time information distributor E will output 9 groups of means and variances according to the input time interval Y, and provide them to 9 time information embedders respectively. The means and variances input to each time information embedder are different.

[0077] 4), The time information embedder fuses the means and variances provided in 3) with the abstract features at the corresponding level, that is, each group of means and variances is respectively conditionally normalized with the input abstract features, so that they present different statistical characteristics matching the time and abstract levels, and the extracted features corresponding to the output are sensitive to time.

[0078] 5), The fusion module at the nth level fuses the two outputs of the time information embedder at this level with the output result of the fusion module at the previous level. The output of the fusion module M9 at the last level of the positive sample pair is used as the positive sample of the discriminator D. The negative sample pair repeats the above processes 3) to 5). The output of the fusion module M 19 of the negative sample pair is used as the negative sample of the discriminator D.

[0079] 6), Train and backpropagate to optimize the dual-channel conditional fusion model through the difference loss function Loss dif and the distribution loss function Loss dis , that is, train the causal feature extraction module and the discriminator D. The causal feature extraction module includes a first-channel module, a second-channel module, and three feature extractors. When the difference loss function Loss dif and the distribution loss function Loss dis converge, the training of the dual-channel conditional fusion model is completed.

[0080] Furthermore, when the training of the entire dual-channel conditional fusion model is completed, the medical image X to be detected is used as the input to the dual-channel conditional fusion model, and the causal rationality score for the medical image X to be detected is output. Calculate the P-value corresponding to the causal rationality score according to the pre-given probability distribution (such as Gaussian distribution). When the P-value is less than 0.05, we determine that this medical image X to be detected does not meet the causal rationality. It should be noted that the P-value can also be adjusted according to actual needs.

[0081] This method introduces a dual-channel conditional fusion model to fuse two causally related images through dual channels and extract their causal features. It can differentially compare and fuse the abstract features from two medical images according to the time interval between the two medical images, and can suppress the differences caused by different patient postures and device parameters, especially the low-level abstract features interfered by patient postures and imaging parameters. By adding a conditional normalization layer, the extracted abstract features are highly sensitive to time, increasing the network's time perception ability. Moreover, through the fusion module, as the training progresses, it can automatically learn to use the features of the required abstraction level, better obtain causal features, overcome the problems in the prior art's causal relationship extraction methods such as the non-adjustable time span and over-sensitivity to the interference caused by different patient postures and imaging parameters, and can obtain more accurate detection results.

[0082] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0083] The above-described embodiments only represent relatively specific and detailed embodiments described in this application, but should not be construed as limiting the scope of the patent application. It should be noted that for those of ordinary skill in the art, without departing from the concept of this application, several modifications and improvements can be made, and these all belong to the protection scope of this application. Therefore, the protection scope of the patent of this application should be subject to the appended claims.

Claims

1. A method for detecting the causal rationality of medical images based on dual-channel conditional fusion, characterized in that: The medical image causal rationality detection method based on dual-channel conditional fusion includes the following steps: S1. Obtain a medical image dataset and select a medical image from the medical image dataset, which is denoted as the original medical image where A is the patient number, i is the shooting and generation date of the original medical image, a represents the shooting part, and b represents the shooting angle; S2. Select two medical images from the medical image dataset using a medical image perturbator, and denote them as the first medical image and the second medical image to make form a positive sample pair that conforms to causal rationality, form a negative sample pair that does not conform to causal rationality. Y is the time interval between the two medical images in the sample pair, B is the patient number, j is the shooting and generation date of the medical image selected by the medical image perturbator, c represents the shooting part, and d represents the shooting angle; The positive sample pair that conforms to causal rationality satisfies the following constraints: Y is the original medical image and the first medical image The time interval, in years; j ≥ i, where both i and j are integers; when i = j, ab ≠ cd; The negative sample pairs that do not conform to causal rationality Meet the following constraints: ①When A = B: Y is the original medical image and the second medical image the time interval between them, in years; i ≥ j, where both i and j are integers; ②When A ≠ B: Y is the original medical image and the second medical image of the time interval, in years; both i and j are integers; S3. Establish a dual-channel conditional fusion model, where the dual-channel conditional fusion model includes a first-channel module, a second-channel module, a discriminator, and three feature extractors. Both the first-channel module and the second-channel module include a time information distributor and multiple sequentially connected embedding fusion units. The embedding fusion unit includes a time information embedder and a fusion module connected in sequence, and performs the following operations: S31. Input the original medical image the first medical image and the second medical image into three feature extractors one by one for feature extraction to obtain the abstract features at each level with corresponding abstract levels from low to high; S32. Input the original medical image and the first medical image into the respective time information embedders of the first channel module one by one in sequence according to the corresponding abstract features of each level from low to high of the abstract level. Input the original medical image and the second medical image into the respective time information embedders of the second channel module one by one in sequence according to the corresponding abstract features of each level from low to high of the abstract level; S33. Use the time information distributor to convert the time interval Y between the two medical images of the sample pair into means and variances matching each level and distribute them to the time information embedders of the corresponding channel modules; S34. Fuse the means and variances received by the embedding fusion units of each channel module into the abstract features matching them to obtain the output features of the corresponding channel modules. The specific operations of each channel module are as follows: S341. Respectively fuse the abstract features of the two medical images received by the time information embedder of the first level with the received means and variances to obtain the corresponding first-level extraction features, and generate first-level fusion features by passing the first-level extraction features through the first-level fusion module; S342. Respectively fuse the abstract features of the two medical images received by the time information embedder of the nth level with the received means and variances to obtain the corresponding nth-level extraction features, and generate nth-level fusion features by passing the nth-level extraction features and the (n - 1)th-level fusion features through the nth-level fusion module, where n ≥ 2; S343. Judge whether n is equal to N, where N is the number of embedding fusion units. If so, it is considered that the fusion of all levels is completed, and the fusion feature of the last level is used as the output feature of the corresponding channel module. Otherwise, set n = n + 1 and return to execute step S342; S4. Sequentially record the output features of the first channel module and the output features of the second channel module as positive samples and negative samples respectively, and input them into a discriminator to obtain the original medical image and obtain the causal rationality score of the 2. The medical image causal rationality detection method based on dual-channel condition fusion according to claim 1, characterized in that: Both the feature extractor and the discriminator adopt multi-layer convolutional neural networks.

3. The method for detecting the causal rationality of medical images based on dual-channel condition fusion according to claim 2, wherein: The feature extractor adopts a ResNet50 network or a Transformer network.

4. The method for detecting the causal rationality of medical images based on dual-channel condition fusion according to claim 1, wherein: The time information distributor adopts a fully connected network.

5. The medical image causal rationality detection method based on dual-channel condition fusion according to claim 1, characterized in that: The time information embedder is a conditional normalization layer.

6. The method for detecting the causal rationality of medical images based on dual-channel condition fusion according to claim 5, wherein: The conditional normalization layer adopts a Batch Normalization layer.

7. The medical image causal rationality detection method based on dual-channel condition fusion according to claim 1, characterized in that: The fusion method of the fusion module is one of multiplying, adding, and taking the maximum value of the pixel-level matrices of its input features.

8. The method for detecting the causal rationality of medical images based on dual-channel condition fusion according to claim 1, wherein: The dual-channel conditional fusion model is also trained and optimized through the difference loss function Loss dif and the distribution loss function Loss dis using backpropagation to optimize the dual-channel conditional fusion model until the difference loss function Loss dif and the distribution loss function Loss dis converges, obtaining a trained dual-channel conditional fusion model.

9. The method for detecting the causal rationality of medical images based on dual-channel condition fusion according to claim 8, wherein: The difference loss function Loss dif and the distribution loss function Loss dis , the formula is as follows: Loss dis = -DD(di(x)) Where x is the positive sample, y is the negative sample, Di(·) is the difference discriminator, is the gradient of Di(x), ||·||1 is the L1 norm, is the square of the L2 norm, and DD(·) is the distribution discriminator.

Images