Image lesion segmentation system based on convolutional neural network

Abstract

The invention discloses an image lesion segmentation system based on a convolutional neural network, and relates to the technical field of image segmentation, and the system comprises a data acquisition module, a cross-modal decoding module, a joint optimization module and a lesion segmentation module. The method is technically characterized by comprising the following steps: performing cross-modal decoding integration on a CT security coding array and an MRI security coding array to generate a fusion lesion segmentation array; performing joint optimization debugging on the CT lesion feature array, the CT segmentation parameter array, the MRI lesion feature array and the MRI segmentation parameter array based on the association relationship between the fusion lesion segmentation array and the model parameter array, and generating optimized segmentation model parameters; using the optimized segmentation model parameters to generate a target lesion segmentation array of the patient set for the CT image and the MRI image; based on the target lesion segmentation array, determining a priority labeling result of the lesion area of the patient; the accuracy of lesion segmentation is improved, and the model calculation efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image segmentation, and in particular to an image lesion segmentation system based on convolutional neural networks. Background Art

[0002] Medical image lesion segmentation is one of the core technologies of computer-aided diagnosis (CAD) systems. It aims to accurately delineate lesion areas (such as tumors, hemorrhages, and tissue abnormalities) from CT and MRI images, providing a quantitative basis for disease diagnosis, treatment planning, and prognosis assessment. Early medical image segmentation techniques primarily relied on manually defined features and thresholds or morphological rules. With the explosive growth of medical imaging data and breakthroughs in deep learning technology, automatic segmentation methods based on convolutional neural networks have gradually replaced traditional techniques and become a research hotspot in this field. In existing image lesion segmentation systems, convolution operations perform well in medical image synthesis tasks. However, convolution operations limit the ability to express contextual features that reflect long-range spatial dependencies, resulting in poor quality medical image synthesis and an inability to well display lesion areas in the synthesized images. Furthermore, due to the inherent shortcomings of CT and MRI image segmentation, for example, CT has limited density resolution for soft tissues and large differences in intensity distribution between different scanning sequences (T1, T2, FLAIR) in MRI images, resulting in poor cross-center data generalization, it is impossible to accurately distinguish between single-modality CT and MRI images, hindering image registration between CT and MRI images and further affecting cross-modality data or image fusion. In addition, during image segmentation, a fixed weight distribution strategy is usually adopted, which cannot dynamically adapt to changes in data quality, resulting in a decrease in overall segmentation accuracy. For example, in brain glioma segmentation, CT is sensitive to calcified areas, and MRI is better for edema zones, but the fixed weights cannot be dynamically adjusted between calcified areas (needing to be dominated by CT) and edema areas (needing to be dominated by MRI). For example, when a certain modality has severe noise (such as low-dose CT) or artifacts (such as MRI motion artifacts), the fixed weight still gives it a fixed contribution, introducing erroneous segmentation. For lesion segmentation results, the traditional influencing lesion segmentation system is not jointly optimized with the model parameters, resulting in a decrease in resource computing and insufficient real-time performance. Summary of the Invention

[0003] (1) Technical problems solved In response to the shortcomings of the existing technology, the present invention provides an image lesion segmentation system based on convolutional neural networks, including a data acquisition module, a cross-modal decoding module, a joint optimization module and a lesion segmentation module. By designing single-modal and cross-modal forms, specific analysis is performed on single CT images, single MRI images, and CT and MRI images to improve the accuracy of data analysis and the accuracy of image registration of CT images and MRI images. During the lesion segmentation process, confidence weights are designed to dynamically adapt to changes in data quality and improve overall segmentation accuracy. The entire system introduces a CNN network to jointly optimize relevant parameters and segmentation model parameters to accelerate model calculation, thereby solving the problems raised in the background technology.

[0004] (2) Technical solution To achieve the above objectives, the present invention is implemented through the following technical solutions: The present application provides an image lesion segmentation system based on a convolutional neural network, the system comprising: A data acquisition module, which acquires a CT safety code array, a CT lesion feature array, and a CT segmentation parameter array, as well as an MRI safety code array, an MRI lesion feature array, and an MRI segmentation parameter array; The cross-modality decoding module performs cross-modality decoding integration on the CT safety coding array and the MRI safety coding array, enhances the CT features and MRI features based on different modalities, and obtains a fused lesion segmentation array; The joint optimization module performs joint optimization and debugging on the CT lesion feature array, CT segmentation parameter array, MRI lesion feature array, and MRI segmentation parameter array based on the association between the fused lesion segmentation array and the model parameter array, and generates optimized segmentation model parameters; The lesion segmentation module uses the optimized segmentation model parameters to generate the target lesion segmentation array for the CT images and MRI images of the patient set; based on the target lesion segmentation array, it determines the priority labeling results of the patient lesion area.

[0005] Furthermore, the CT security coding array is a security coding array of the patient set for the lesion segmentation array of the CT image; The CT lesion feature array and the CT segmentation parameter array are model parameter arrays of the CT lesion segmentation array; The MRI security coding array is a security coding array of the lesion segmentation array of the patient set for the MRI image; The MRI lesion feature array and the MRI segmentation parameter array are model parameter arrays of the MRI lesion segmentation array.

[0006] Furthermore, the steps of cross-modal decoding integration include: Decrypt the CT security code array and output the CT image; Decrypt the MRI security code array and output the MRI image; Determine whether it is a cross-modal form: When the decryption result contains only CT images or MRI images, it is determined to be a single-modality form; When the decryption result includes CT images and MRI images, it is determined to be cross-modal, and non-rigid registration is performed on the CT images and MRI images, and the CT features and MRI features are fused through channel cascade and attention mechanism; The characteristic variance and modal contribution ratio of each modality are calculated, the corresponding segmentation branch is dynamically selected, and the corresponding fusion lesion segmentation array is generated; wherein the segmentation branch includes at least one of a straight line segment branch, an oblique line segment branch, a curved line segment branch, a highly complex branch, and a basic branch.

[0007] Furthermore, the step of determining whether it is a single modal form includes: CT images only: CT energy spectrum curves are constructed based on CT images. The local maximum method is used to identify peak regions in the CT energy spectrum curves. The residual block structure is activated in the peak regions. The corresponding energy values, intensity values, and penetration depths are extracted based on the peak regions. A two-dimensional graph of energy value / intensity value and penetration depth / intensity value is created, and the line segment type is determined. Feature fusion is then performed in the residual block structure, and the fusion weight is automatically adjusted based on the proportion of line segment types. Line segment types include straight segments, oblique segments, and curved segments. MRI images only: Extract multi-sequence MRI data based on MRI images, including T1WI, T2WI, and QSM; The edema area was extracted based on T2WI and an edema mask was generated; Extracting metal deposition areas based on QSM, including at least one of copper deposition areas and iron deposition areas; Combined with T1WI, a deposition mask is generated; When there is spatial overlap between the edema area and the metal deposition area, or the proportion of the metal deposition area exceeds a preset threshold, the high-complexity branch is triggered, using T1WI and T2WI as query vectors, and the deposition mask and edema mask as key values ​​to obtain the attention weight matrix to enhance the MRI features; otherwise, the basic branch is triggered; At the same time, the bottleneck block structure is activated to perform channel compression and expansion on the outputs of highly complex branches and basic branches.

[0008] Furthermore, the association relationship is achieved through at least one of the following methods: The strength of the association between the fused lesion segmentation array and the model parameter array was quantified based on the Pearson correlation coefficient or Spearman rank correlation coefficient between the two.

[0009] Furthermore, the steps of joint optimization and debugging include: Based on the association between the CT lesion feature array, the CT segmentation parameter array, the MRI lesion feature array, the MRI segmentation parameter array, and the fusion lesion segmentation array, the Nth round of parameter debugging is performed; Perform lesion inference on the CT lesion feature array, CT segmentation parameter array, MRI lesion feature array, and MRI segmentation parameter array after the Nth round of debugging; Obtaining an Nth segmentation error value based on the Nth CT lesion inference array, the Nth MRI lesion inference array, and a preset error calculation function; When the absolute difference between the Nth segmentation error value and the Mth segmentation error value is less than the preset threshold, the segmentation model parameters after the Nth round of debugging are determined as the final optimization result; otherwise, set M=N, N=N+1, and return to the iteration; Wherein, N is a natural number greater than 0, and each round of debugging uses a gradient descent algorithm to update the CT segmentation parameters or MRI segmentation parameters.

[0010] Furthermore, the steps of lesion reasoning include: Load the model architecture optimized after the Nth round of debugging and select the inference form, including single-modal reasoning and cross-modal reasoning; Based on the inference form, the CT lesion feature array, CT segmentation parameter array, MRI lesion feature array and MRI segmentation parameter array are used as model condition inputs, and the predicted lesion probability is output.

[0011] Furthermore, the algorithm model of the error calculation function is: In the single-mode form, the Dice coefficient error function is used; In the cross-modal form, the weighted joint Dice loss function is used.

[0012] Furthermore, the step of prioritizing the result includes: Loading the optimized segmentation model parameters, obtaining confidence weights for the CT lesion segmentation array and the MRI lesion segmentation array, wherein the confidence weights are based on at least one of probability map entropy or cross-modality overlap area ratio; Based on the confidence weight, weighted sum is performed to obtain the score evaluation coefficient Priority; The score evaluation coefficient and the standard score interval [DF min , DF max Comparative Analysis: When Priority≤DF min , mark the Priority as a, edit it as a first-level character, and combine a and the first-level character to generate a low-priority lesion area; When DF min <Priority≤DF max, mark the Priority as b, edit it as a secondary character, and combine b and the secondary character to generate a medium-priority lesion area; When Priority>DF max , mark the Priority as c, edit it as a third-level character, and combine c and the third-level character to generate a high-priority lesion area.

[0013] (3) Beneficial effects The present invention provides an image lesion segmentation system based on convolutional neural network, which has the following beneficial effects: 1. Through decoding operations, the present invention directly accesses the original eigenvalues, restores the spatial attributes, numerical continuity, and metadata integrity of CT and MRI images, supports statistical modeling and machine learning, and makes the coordinate alignment, feature association, and model reasoning required for cross-modal fusion possible. This facilitates the subsequent analysis of the mutual fusion of CT and MRI, and improves diagnostic accuracy and reliability. 2. The present invention designs two different modal forms, including a single-modal form and a cross-modal form. In the single-modal form, for CT images, a two-dimensional graph is constructed with energy value / intensity value as the X-axis and penetration depth / intensity value as the Y-axis. The line segment type is determined based on the two-dimensional graph, and the residual block structure is activated to enhance CT features according to different line segment types. For MRI images, edema areas and metal deposition areas are extracted, and the bottleneck block structure is activated to perform channel compression and expansion on the outputs of highly complex branches and basic branches. In the cross-modal form, non-rigid registration and fusion of CT and MRI images are used to solve the registration offset problem caused by resolution and contrast differences between CT and MRI, thereby improving the accuracy of lesion segmentation. 3. This invention significantly improves the accuracy and clinical practicality of medical image segmentation by setting confidence weights and dynamically adjusting the contribution of each modality. It also adapts data quality and enhances the credibility of results, providing more reliable decision support for AI-assisted diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a module diagram of an image lesion segmentation system according to an exemplary embodiment; Figure 2 is a module block diagram in a single-mode form according to an exemplary embodiment; Figure 3 This is a flowchart of the non-rigid registration and fusion of CT and MRI. DETAILED DESCRIPTION

[0015] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0016] Example The embodiment of the present invention provides an image lesion segmentation system based on convolutional neural network; Figure 1 is a module diagram of an image lesion segmentation system according to an exemplary embodiment; Figure 2 is a module block diagram in a single-mode form according to an exemplary embodiment; Figure 3 This is a flowchart for the non-rigid registration and fusion of CT and MRI; see Figure 1 - Figure 3 The system includes: a data acquisition module, a cross-modal decoding module, a joint optimization module, and a lesion segmentation module, and the data acquisition module, the cross-modal decoding module, the joint optimization module, and the lesion segmentation module are communicatively connected; Data acquisition module: acquires the CT security code array, CT lesion feature array, and CT segmentation parameter array uploaded by the CT imaging platform, as well as the MRI security code array, MRI lesion feature array, and MRI segmentation parameter array uploaded by the MRI imaging platform; The CT security coding array is a security coding array of the lesion segmentation array of the patient set for the CT image; The CT lesion feature array and the CT segmentation parameter array are model parameter arrays of the CT lesion segmentation array; The MRI security coding array is a security coding array of the lesion segmentation array of the patient set for the MRI image; The MRI lesion feature array and the MRI segmentation parameter array are model parameter arrays of the MRI lesion segmentation array; The following is an explanation of the nouns involved: Secure code array: a code array consisting of encrypted and anonymized image data; Lesion feature array: deep features extracted by convolutional neural network (CNN), that is, a set of quantitative features extracted from lesion segmentation results; for example, a set of quantitative features extracted from CT lesion segmentation results; Segmentation parameter array: A set of configuration parameters required for accurately locating lesion areas. For example, this array can be used to annotate abnormal tissue areas (such as deposits or edema) pixel by pixel in CT or MRI images, and to generate binary masks or probabilities, along with the corresponding trainable parameters or hyperparameters (such as learning rate and number of iterations). The lesion feature array and segmentation parameter array together constitute the model input or configuration parameters used to perform lesion segmentation tasks in medical images. Coordinated optimization of the two (such as automated parameter adjustment) is key to improving the accuracy of medical image analysis, especially in clinical diagnosis and scientific research. They must be clearly distinguished and applied appropriately. Lesion segmentation array: It is the lesion segmentation result output by the system after executing the segmentation process, and the feature set is quantified; The security code array, lesion feature array, and segmentation parameter array are all uploaded from the imaging platform. Together, these three arrays form a secure, accurate, and scalable medical image analysis system. This system aims to optimize collaboration between different modalities (e.g., single-modality CT, single-modality MRI, and cross-modality CT and MRI), ensuring consistency and adaptability of lesion segmentation results. This system is suitable for scenarios such as tumor screening, emergency diagnosis, and surgical planning. The cross-modality decoding module performs cross-modality decoding integration on the CT safety coding array and the MRI safety coding array, enhances the CT features and MRI features based on different modalities, and obtains a fused lesion segmentation array; The steps of cross-modal decoding integration include: Decrypt the CT security code array and output the CT image; Decrypt the MRI security code array and output the MRI image; For example, when uploading data to a CT imaging platform, the platform encrypts the collected CT images and their patient data to generate a secure CT encoding array. For example, differential privacy processing is performed on the pixel coordinates of the CT image, controllable noise is added, or homomorphic encryption is used for the lesion location. The specific choice depends on security requirements and performance considerations. The encrypted array is uploaded to the computer system in binary or Base64 encoding format to ensure data security during transmission and storage. Through encryption operations, patient collections and their imaging data (CT images, MRI images) are traceable. The above data are usually stored in the form of binary ciphertext or hash values ​​in the blockchain to associate with the hashed patient ID. At the same time, role-based (RBAC) control decryption permissions are set (for example, only radiologists can decrypt CT), forming a chain structure to achieve dynamic authorization; through decoding operations, the original feature values ​​are directly accessed, and the spatial attributes, numerical continuity and metadata integrity of CT images and MRI images are restored, supporting statistical modeling and machine learning, making the coordinate alignment, feature association and model reasoning required for cross-modal fusion possible, facilitating the subsequent analysis of the mutual fusion of CT and MRI, and improving diagnostic accuracy and reliability; in addition, the decoded data needs to be re-encrypted and stored, and the access log retention time is not less than the set data life cycle, and it will be automatically deleted when the preset data life cycle is exceeded; Determine whether it is a cross-modal form: When the decryption result contains only CT images or MRI images, it is determined to be a single-modality form; conversely, when the decryption result contains both CT images and MRI images, it is determined to be a cross-modality form; When in single mode: CT images only: A CT energy spectrum curve is constructed based on CT images. The local maximum method is used to identify the peak area in the CT energy spectrum curve. The residual block structure is activated in the peak area to force the system to have higher segmentation accuracy in the peak area. Specifically, the steps of the local maximum method include: The CT energy spectrum curve describes how the attenuation characteristics of X-rays change with photon energy when they pass through matter. It can be used to optimize the X-ray source energy spectrum curve for CT scanning parameters, or it can be used as a material attenuation energy spectrum curve. The CT energy spectrum curve is used to analyze material composition, distinguish different tissues or contrast agents (such as iodine and calcium), etc. The attenuation coefficient of each X-ray photon energy point is scanned in the CT energy spectrum curve, and the point that is Δμ higher than the neighboring point is marked as the peak; Δμ represents the peak amplitude, which is expressed as the difference between the peak point value and the baseline; the pseudo-peaks caused by noise are removed by threshold filtering, and the valid peak points are retained. The corresponding peak area is determined based on the valid peak points; in addition, the CT energy spectrum curve may contain noise, so the CT energy spectrum curve needs to be smoothed, such as using Gaussian filtering, median filtering, or other filtering methods to remove noise; Extract the corresponding energy value, intensity value and penetration depth based on the peak area, draw a two-dimensional graph with the energy value / intensity value as the X-axis and the penetration depth / intensity value as the Y-axis, and determine the line segment type based on the two-dimensional graph, including straight line segment, oblique line segment and curved line segment; The steps to determine the line segment type are as follows: When at least one of a horizontal line or a vertical line exists, it is determined to be a straight line segment, and the CT features are uniformly enhanced using a two-point algorithm: Used to suppress redundant information and improve computing efficiency; Where α(E, d) represents the gain function, which is used to uniformly enhance CT features, E represents the energy value, E0 represents the center energy, ΔE represents the energy value range, and α base1 and α base2 Both represent the reference gain, and α base1 >α base2 >0, d represents the penetration depth, L0 represents the horizontal line, and L1 represents the vertical line; Formula explanation: When it is a horizontal line, the penetration depth contrast is stretched based on the energy value range; when it is a vertical line, the penetration depth is introduced, the energy response is enhanced, metal artifacts are suppressed, and the soft tissue visualization is enhanced. The combination of horizontal and vertical lines enables accurate visualization of CT images, providing high-value imaging support for clinical diagnosis and treatment planning. When there is at least one of a positive or negative slope straight line, it is determined to be a slope segment, and the bilinear interpolation algorithm is used to linearly enhance the CT features to strengthen the structure boundary and suppress noise; For example, we extract the inflection point position from the oblique line segment and define the weight matrix at the inflection point. We then apply bilinear interpolation to each sub-segment to ensure a smooth transition at the connection, merge the enhancement effects of each sub-segment, and eliminate boundary mutations. When at least one of the exponential and polynomial expressions exists, it is determined to be a curve segment, and the preset nonlinear model algorithm is used to enhance the complex CT features, which is used to capture nonlinear features and enhance the model's ability to discriminate complex pathologies; For example, at the curve segment, Sigmoid or polynomial functions are used to enhance high-contrast areas (such as calcification) and enhance high-frequency details of the curve segment; The residual block structure includes straight line segment branches, oblique line segment branches, and curved line segment branches. By determining the line segment type and capturing linear and nonlinear relationships, it can effectively extract deep features of the CT energy spectrum. Feature fusion is then performed in the residual block structure: , and the fusion weight is automatically adjusted based on the proportion of line segment types; where F fuse Represents the fused CT features, i represents the type of line segment, including straight line segment, oblique line segment and curved line segment, represents the CT features obtained under the i-th line segment type, λ represents the weight, including uniform enhancement weight, linear enhancement weight and nonlinear enhancement weight, and the uniform enhancement weight, linear enhancement weight and nonlinear enhancement weight correspond to the straight line segment, oblique line segment and curved line segment respectively, and λ i =1; For example: if the curve segment accounts for 60%, the nonlinear enhancement weight is 0.6; MRI images only: Extract multi-sequence MRI data based on MRI images, including T1WI, T2WI and QSM; Combined with T2WI, edema area was extracted and edema mask was generated; Specifically, the edema area can be identified by the signal intensity threshold segmentation method, because the edema area has increased water content in the tissue, resulting in a prolonged T2 relaxation time and a high signal on T2WI. The above steps are not described in detail. Combined with QSM, metal deposition areas are extracted, including at least one of copper deposition areas and iron deposition areas; Combined with T1WI, a deposition mask is generated; When there is spatial overlap between the edema area and the metal deposition area, or the proportion of the metal deposition area exceeds a preset threshold, for example, the T1WI signal intensity in the copper deposition area is higher than the preset percentile (generally 10% of voxels), the high-complexity branch is triggered. T1WI and T2WI are used as query vectors, and the deposition mask and edema mask are used as key values ​​to obtain the attention weight matrix to enhance the MRI features. Otherwise, the base branch is triggered; Activate the bottleneck block structure to perform channel compression and expansion on the outputs of highly complex branches and basic branches; Specifically, metal deposition (such as iron deposition, calcification or hemorrhage residue in the brain) is manifested as a significant increase in local magnetic susceptibility in QSM (quantitative magnetic susceptibility imaging). The combined analysis of QSM and T1WI can significantly improve the specificity of metal deposition areas, especially in distinguishing different pathological types such as calcification, iron deposition and hemorrhage. Therefore, the metal deposition area is the basis of imaging lesion analysis. For example, the cause of Wilson's disease is: ATP7B gene mutation leads to copper metabolism disorder, copper is deposited in the liver, cornea and basal ganglia; and the cause of Parkinson's disease is: degeneration of dopaminergic neurons in the substantia nigra, accompanied by iron ions (Fe 3+ ) Abnormal deposition; by analyzing the metal deposition area, the image lesion is segmented more finely; at the same time, the edema area and the metal deposition area are given special consideration, which enables deeper extraction of MRI features; When in cross-modal form: non-rigid registration of CT images and MRI images is performed, and CT features and MRI features are fused through channel cascade and attention mechanism; Specifically, non-rigid registration includes using the Demons algorithm or B-spline free deformation, optimizing mutual information as a similarity measure, balancing the segmentation errors of CT and MRI through multiple rounds of debugging, and avoiding overfitting of the model to a single modality; based on the CT images and MRI images to be registered, a large number of data sets are formed and divided into test sets, training sets, and validation sets; during the training process, a dual UNet network is used to extract deep features of CT images and MRI images through CT feature extractors and MRI feature extractors respectively; CT features and MRI features are channel-concatenated to generate cascade features; the cascade features are dynamically weighted and fused in spatial and channel dimensions through the attention fusion module to generate fusion features; during the training process, the model weights of each epoch are saved, the validation set is inferred, and a lesion probability map and evaluation indicators are generated. After each epoch of training, the model evaluation of the validation set is performed, for example: the model weight with the highest Dice coefficient in the validation set is selected. If the Dice coefficient is the same, the model with a smaller HD95 is preferred. Calculate the characteristic variance of each modality and the corresponding modality contribution ratio, dynamically select the corresponding segmentation branch, and generate the corresponding fusion lesion segmentation array; wherein the segmentation branch includes at least one of a straight line segment branch, an oblique line segment branch, a curved line segment branch, a highly complex branch, and a basic branch; The joint optimization module performs joint optimization and debugging on the CT lesion feature array, CT segmentation parameter array, MRI lesion feature array, and MRI segmentation parameter array based on the association between the fused lesion segmentation array and the model parameter array, and generates optimized segmentation model parameters; The steps of joint optimization and debugging include: Specifically, the image lesion segmentation system is based on a convolutional neural network (CNN) architecture and includes at least a generator and a discriminator. The generator consists of an encoder, a bottleneck layer, and a decoder. The encoder and decoder contain several residual block structures or bottleneck block structures to extract and reconstruct metal deposition features, edema features, etc. in the image. In addition, a generative adversarial network (GAN) is also set up on the CNN network to balance the difference in segmentation accuracy between CT and MRI. The discriminator network distinguishes the segmentation result features of CT and MRI and obtains the CT segmentation parameter array and MRI segmentation parameter array respectively, as well as the corresponding fused lesion segmentation array. Based on the association between the CT lesion feature array, the CT segmentation parameter array, the MRI lesion feature array, the MRI segmentation parameter array, and the corresponding fusion lesion segmentation array, the Nth round of parameter debugging is performed to obtain: CT lesion feature array after N rounds of debugging, CT segmentation parameter array after N rounds of debugging, MRI lesion feature array after N rounds of debugging, MRI segmentation parameter array after N rounds of debugging; Where N is a natural number greater than 0, and each round of debugging uses the gradient descent algorithm to update the CT segmentation parameters or MRI segmentation parameters, and aligns the lesion feature distributions of CT and MRI through feature space projection; The association relationship is achieved through at least one of the following methods: Obtain the Pearson correlation coefficient or Spearman rank correlation coefficient between the fusion lesion segmentation array and the model parameter array to quantify the correlation between the two and reflect the strength of the association; Specifically, the Pearson correlation coefficient measures the linear relationship between the two and is applicable to normally distributed data (e.g., volume features, regular shape features); the Spearman rank correlation coefficient measures the monotonic relationship and is applicable to non-normal or nonlinear data (e.g., texture features, hyperparameters). The specific steps are not described in detail here. Perform lesion inference on the CT lesion feature array, CT segmentation parameter array, MRI lesion feature array, and MRI segmentation parameter array after the Nth round of debugging to obtain: Nth CT lesion inference array; Nth MRI lesion inference array; The steps of lesion inference include: Load the model architecture optimized after the Nth round of debugging and select the inference form, including single-modal reasoning and cross-modal reasoning; Based on the inference form, the CT lesion feature array, CT segmentation parameter array, MRI lesion feature array and MRI segmentation parameter array are used as model condition inputs, and the predicted lesion probability is output; Unimodal reasoning: The CT lesion feature array and the CT segmentation parameter array are used as model condition inputs. For example, the model architecture selects a 3DU-Net encoder and outputs the CT lesion probability. The MRI lesion feature array and the MRI segmentation parameter array are used as model condition inputs. For example, the model architecture selects the Swin-Transformer encoder to output the MRI lesion probability. Cross-modal reasoning: CT lesion feature array, CT segmentation parameter array, MRI lesion feature array, MRI segmentation parameter array, and fused lesion segmentation array are used as model condition inputs to output the predicted lesion probability; Obtaining an Nth segmentation error value based on the Nth CT lesion inference array, the Nth MRI lesion inference array, and a preset error calculation function; When the absolute difference between the Nth segmentation error value and the Mth segmentation error value is less than the preset threshold, the segmentation model parameters after the Nth round of debugging are determined as the final optimization result; otherwise, set M=N, N=N+1, and return to the iteration; The Nth segmentation error value includes the Nth CT segmentation error value corresponding to the Nth CT lesion inference array and the Nth MRI segmentation error value corresponding to the Nth MRI lesion inference array; when the absolute difference between the Nth segmentation error value and the Mth segmentation error value is less than a preset threshold, it means that the absolute value of the Nth CT segmentation error value and the Mth CT segmentation error value is less than the preset threshold, and the absolute value of the Nth MRI segmentation error value and the Mth MRI segmentation error value is less than the preset threshold; The algorithm model of the preset error calculation function is: In the single-mode form, the Dice coefficient error function is used; In the cross-modal form, the weighted joint Dice loss function is used; The following is an explanation of the nouns involved: Fusion lesion segmentation array: a comprehensive data matrix generated by the segmentation results of medical images in various modalities (e.g., single-modality CT, single-modality MRI, cross-modality CT and MRI); Lesion inference array, a set of parameters used by the system to make decisions about lesion regions, based on high-level semantic information generated by the lesion segmentation array and cross-modal data, and a quantized feature set; Segmentation model parameters: refers to the weights, biases, and convolution kernels within the trained model; The lesion segmentation module uses the optimized segmentation model parameters to generate target lesion segmentation arrays for CT and MRI images of the patient set; based on the target lesion segmentation arrays, it determines the priority labeling results of the patient lesion areas; The step of determining the priority marking result includes: Loading the optimized segmentation model parameters, obtaining confidence weights for the CT lesion segmentation array and the MRI lesion segmentation array, wherein the confidence weights are based on at least one of probability map entropy or cross-modality overlap area ratio; By formula: Get the score evaluation coefficient Priority; In the formula, β1 and β2 are adjustment coefficients, and β1+β2=1. represents the confidence weight based on the ratio of cross-modal overlapping areas, represents the confidence weight based on the probability graph entropy, represents the normalized probability graph entropy, J represents the total number of modes, and its value is 1 or 2. When it is a single-modal form, the value is 1, and when it is a cross-modal form, the value is 2; The score evaluation coefficient and the standard score interval [DF min , DF max Comparative Analysis: When Priority≤DF min , mark the Priority as a, edit it as a first-level character, and combine a and the first-level character to generate a low-priority lesion area; When DF min <Priority≤DF max , mark the Priority as b, edit it as a secondary character, and combine b and the secondary character to generate a medium-priority lesion area; When Priority>DF max , mark the Priority as c, edit it as a third-level character, and combine c and the third-level character to generate a high-priority lesion area; For example: For patient XX, obtain CT and MRI images of the slope area; Based on CT images, a CT energy spectrum curve is constructed. CT features are enhanced based on various line segment types to identify tumor calcification morphology (speckled, ring-shaped, or arcuate), bone destruction type (osteolytic, expansive, ground-glass), and adjacent bone changes (scalloped indentations, sclerotic edges). For example, if only vertical lines are present, soft tissue visualization is enhanced, allowing for clearer identification of calcification morphology. Based on MRI images, the system extracts areas of edema and metal deposition; edema within 3 mm of calcification indicates an active tumor. If a cross-modality pattern is identified, CT and MRI features are fused to further enhance the features and generate a fused lesion segmentation array. The following enhanced feature descriptions are listed in Table 1: Table 1 Enhanced feature description data

[0017] Low-priority lesion areas, medium-priority lesion areas, and high-priority lesion areas are transmitted in real time to the medical picture archiving and communication system (PACS) as priority annotation results, and overlaid on the visual interface of the doctor's workstation. The annotation results support export in the DICOM standard format. In addition, the priority annotation results are consistent with the doctor's diagnostic logic. For example, for medium-priority lesion areas, a multidisciplinary consultation is required within 72 hours to formulate a surgical / radiotherapy plan. Confidence-driven weights dynamically adjust the contribution of each modality by quantifying single-modal uncertainty and cross-modal consistency, solving the problems of rigid weights and reliance on prior knowledge in traditional methods, and significantly improving the accuracy and clinical practicality of medical image segmentation. Its core value lies in adaptive data quality and enhanced result credibility, providing more reliable decision support for AI-assisted diagnosis.

[0018] In the application, the several formulas involved are all calculated by taking their numerical values ​​after removing the dimensions, and the formula is a formula of the most recent real situation obtained by collecting a large amount of data and performing software simulation. The formula is set by technical personnel in this field according to actual conditions.

[0019] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0020] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0021] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. Image lesion segmentation system based on convolutional neural network, characterized by: The system comprises: A data acquisition module, which acquires a CT safety code array, a CT lesion feature array, and a CT segmentation parameter array, as well as an MRI safety code array, an MRI lesion feature array, and an MRI segmentation parameter array; The cross-modality decoding module performs cross-modality decoding integration on the CT safety coding array and the MRI safety coding array, enhances the CT features and MRI features based on different modalities, and obtains a fused lesion segmentation array; The joint optimization module performs joint optimization and debugging on the CT lesion feature array, CT segmentation parameter array, MRI lesion feature array, and MRI segmentation parameter array based on the association between the fused lesion segmentation array and the model parameter array, and generates optimized segmentation model parameters; The lesion segmentation module uses the optimized segmentation model parameters to generate the target lesion segmentation array for the CT images and MRI images of the patient set; based on the target lesion segmentation array, it determines the priority labeling results of the patient lesion area.

2. The convolutional neural network-based image lesion segmentation system according to claim 1, characterized in that: The CT security coding array is a security coding array of the lesion segmentation array of the patient set for the CT image; The CT lesion feature array and the CT segmentation parameter array are model parameter arrays of the CT lesion segmentation array; The MRI security coding array is a security coding array of the lesion segmentation array of the patient set for the MRI image; The MRI lesion feature array and the MRI segmentation parameter array are model parameter arrays of the MRI lesion segmentation array.

3. The image lesion segmentation system based on convolutional neural network according to claim 1, characterized in that: The steps of cross-modal decoding integration include: Decrypt the CT security code array and output the CT image; Decrypt the MRI security code array and output the MRI image; Determine whether it is a cross-modal form: When the decryption result contains only CT images or MRI images, it is determined to be a single-modality form; When the decryption result includes CT images and MRI images, it is determined to be cross-modal, and non-rigid registration is performed on the CT images and MRI images, and the CT features and MRI features are fused through channel cascade and attention mechanism; The characteristic variance and modal contribution ratio of each modality are calculated, the corresponding segmentation branch is dynamically selected, and the corresponding fusion lesion segmentation array is generated; wherein the segmentation branch includes at least one of a straight line segment branch, an oblique line segment branch, a curved line segment branch, a highly complex branch, and a basic branch.

4. The image lesion segmentation system based on convolutional neural network according to claim 3, characterized in that: The steps to determine whether it is a single modal form include: CT images only: CT energy spectrum curves are constructed based on CT images. The local maximum method is used to identify peak regions in the CT energy spectrum curves. The residual block structure is activated in the peak regions. The corresponding energy values, intensity values, and penetration depths are extracted based on the peak regions. A two-dimensional energy / intensity-penetration depth / intensity map is created. The line segment type is determined, and feature fusion is performed in the residual block structure. The fusion weight is automatically adjusted based on the proportion of line segment types. Line segment types include straight segments, oblique segments, and curved segments. MRI images only: Extract multi-sequence MRI data based on MRI images, including T1WI, T2WI, and QSM; The edema area was extracted based on T2WI and an edema mask was generated; Extracting metal deposition areas based on QSM, including at least one of copper deposition areas and iron deposition areas; Combined with T1WI, a deposition mask is generated; When there is spatial overlap between the edema area and the metal deposition area, or the proportion of the metal deposition area exceeds a preset threshold, the high-complexity branch is triggered, using T1WI and T2WI as query vectors, and the deposition mask and edema mask as key values ​​to obtain the attention weight matrix to enhance the MRI features; otherwise, the basic branch is triggered; At the same time, the bottleneck block structure is activated to perform channel compression and expansion on the outputs of highly complex branches and basic branches.

5. The image lesion segmentation system based on convolutional neural network according to claim 1, characterized in that: The association relationship is achieved through at least one of the following methods: The strength of the association between the fused lesion segmentation array and the model parameter array was quantified based on the Pearson correlation coefficient or Spearman rank correlation coefficient between the two.

6. The image lesion segmentation system based on convolutional neural network according to claim 1, characterized in that: The steps of joint optimization and debugging include: Based on the association between the CT lesion feature array, the CT segmentation parameter array, the MRI lesion feature array, the MRI segmentation parameter array, and the fusion lesion segmentation array, the Nth round of parameter debugging is performed; Perform lesion inference on the CT lesion feature array, CT segmentation parameter array, MRI lesion feature array, and MRI segmentation parameter array after the Nth round of debugging; Obtaining an Nth segmentation error value based on the Nth CT lesion inference array, the Nth MRI lesion inference array, and a preset error calculation function; When the absolute difference between the Nth segmentation error value and the Mth segmentation error value is less than the preset threshold, the segmentation model parameters after the Nth round of debugging are determined as the final optimization result; otherwise, set M=N, N=N+1, and return to the iteration; Wherein, N is a natural number greater than 0, and each round of debugging uses a gradient descent algorithm to update the CT segmentation parameters or MRI segmentation parameters.

7. The image lesion segmentation system based on convolutional neural network according to claim 6, characterized in that: The steps of lesion inference include: Load the model architecture optimized after the Nth round of debugging and select the inference form, including single-modal reasoning and cross-modal reasoning; Based on the inference form, the CT lesion feature array, CT segmentation parameter array, MRI lesion feature array and MRI segmentation parameter array are used as model condition inputs, and the predicted lesion probability is output.

8. The image lesion segmentation system based on convolutional neural network according to claim 6, characterized in that: The algorithm model of the error calculation function is: In the single-mode form, the Dice coefficient error function is used; In the cross-modal form, the weighted joint Dice loss function is used.

9. The image lesion segmentation system based on convolutional neural network according to claim 1, characterized in that: Steps to prioritize results include: Loading the optimized segmentation model parameters, obtaining confidence weights for the CT lesion segmentation array and the MRI lesion segmentation array, wherein the confidence weights are based on at least one of probability map entropy or cross-modality overlap area ratio; Based on the confidence weight, weighted sum is performed to obtain the score evaluation coefficient Priority; The score evaluation coefficient and the standard score interval [DF min , DF max Comparative Analysis: When Priority≤DF min , mark the Priority as a, edit it as a first-level character, and combine a and the first-level character to generate a low-priority lesion area; When DF min <Priority≤DF max , mark the Priority as b, edit it as a secondary character, and combine b and the secondary character to generate a medium-priority lesion area; When Priority>DF max , mark the Priority as c, edit it as a third-level character, and combine c and the third-level character to generate a high-priority lesion area.

Images