Combined brain tumor MRI classification method based on domain alignment evidence

Abstract

The invention belongs to the field of brain tumor classification, and particularly relates to a combined brain tumor MRI classification method based on domain alignment evidence, and the method comprises the steps of pre-defining, framework integration, domain alignment, evidence guide fusion, uncertainty weighted estimation based on fusion, uncertainty image weight reduction and tumor sub-region output. According to the scheme, a feature interaction bridge is established through two-way cross attention to achieve multi-scale two-way information complementation, shared embedding is generated, statistical attributes of a source domain and a target domain are matched, feature distribution convergence is forced by executing an adversarial alignment operation, a tumor core and a fuzzy boundary are effectively described, and missing detection and false positive are reduced; according to the method, difference priori is obtained through the difference value of the original tumor MRI image and the reconstructed tumor-free contrast image, the weight of double-source evidence is reduced by using a pixel-level uncertainty map, accurate and complete positioning guidance is provided for a segmentation decoder, and the positioning accuracy and boundary integrity of a segmentation task are improved.

Description

Technical Field

[0001] This invention relates to the field of brain tumor classification, specifically to a combined MRI classification method for brain tumors based on domain alignment evidence. Background Technology

[0002] The joint brain tumor MRI classification method based on domain alignment evidence refers to a method for classifying and segmenting brain tumors under conditions of few samples and weak supervision.

[0003] Among existing approximate solutions, such as CN119741312B, "A Deep Learning-Based MRI Medical Image Segmentation Method and System," this solution addresses the technical problems of low tumor identification accuracy and insufficient classification precision caused by significant interference from non-interest areas such as background noise and subcutaneous fat in breast MRI images, as well as blurred details and unclear edges in lesion areas. It employs standardized preprocessing, localization and detail enhancement of suspected lesion areas, three-channel feature enhancement, deep learning recognition, and multi-level SVM classification. First, non-interest areas are removed through dynamic threshold segmentation and morphological operations. Then, a saliency map is generated based on grayscale gradients to locate abnormal sides, followed by Gaussian modeling. The preprocessed image is decomposed and dynamically enhanced and reconstructed. Then, a three-channel input is generated through noise reduction, contrast enhancement, and edge enhancement. The input is used to a deep learning model with residual module and SE channel attention mechanism to complete tumor identification. Finally, multi-dimensional features are extracted and a two-level SVM classifier is used to determine benignity, malignancy and subtype. This achieves accurate preprocessing of breast MRI images, efficient tumor identification and accurate classification, and improves detection accuracy and robustness. It provides reliable auxiliary diagnosis for clinical use. However, there are technical problems such as domain offset of cross-device data in MRI segmentation and lack of pixel-level annotation in weak supervision scenarios, which leads to poor model generalization ability, blurred segmentation boundaries and easy to miss small lesions.

[0004] Furthermore, for example, CN118262105B describes a segmentation method and apparatus for multimodal tumor images based on fully supervised contrastive learning. This approach addresses the challenges of existing segmentation models that only optimize the network architecture while neglecting feature space mining, leading to low segmentation accuracy in difficult-to-segment regions and insufficient voxel classification accuracy. It employs a combination of fully supervised contrastive learning, anchor point sampling, and local positive and negative sample selection: using dResUNet as the basic architecture, a PCM module is introduced into the decoder, and the segmentation results are predicted from sample images. By comparing the results with the true labels to identify incorrect and correct voxels, and sampling difficult-to-segment anchor points based on the segmentation window confidence, positive and negative sample voxels with the best confidence are selected within the neighborhood of the anchor points. The model is trained by combining InfoNCE contrast loss, cross-entropy loss, and Dice loss, which achieves accurate segmentation of tumor regions in multimodal tumor images and significantly improves the segmentation accuracy of difficult-to-segment regions. The Dice index of each region reaches the advanced level. However, there are technical problems in traditional segmentation models, such as the lack of specificity in the fusion of multi-source weakly supervised evidence and the large interference from unreliable regions, which leads to insufficient effectiveness of prior guidance. Summary of the Invention

[0005] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides a joint brain tumor MRI classification method based on domain-aligned evidence. Addressing the technical problems in brain tumor MRI segmentation, such as domain shift in cross-device data and lack of pixel-level annotations in weakly supervised scenarios, which lead to poor model generalization, blurred segmentation boundaries, and easy omission of small lesions, this solution employs a lightweight ViT backbone network to construct a dual-view encoder. A feature interaction bridge is established through bidirectional cross-attention to achieve multi-scale bidirectional information complementarity and generate shared embeddings. Domain alignment is performed to match the statistical attributes of the source and target domains. Adversarial alignment is then performed to force feature distribution convergence. Evidence-guided fusion generates heatmaps, difference priors, and pixel-level uncertainty maps. Uncertainty-weighted estimation transforms multi-source weakly supervised evidence into a continuous prior probability map. Combined with mask filtering to ignore unreliable pixels, a lightweight U-Net segmentation decoder fuses features and the prior map to perform segmentation, achieving a dual improvement in cross-domain robustness and weakly supervised segmentation accuracy. This approach effectively characterizes the tumor core and ambiguous boundaries, reducing missed detections and false positives. Addressing the technical issues of insufficient prior guidance effectiveness due to the lack of targeted multi-source weakly supervised evidence fusion and significant interference from unreliable regions in traditional segmentation models, this solution employs an evidence-guided fusion process. It generates a heatmap based on classification features as initial localization evidence, obtains a difference prior by comparing the original tumor MRI image with a reconstructed tumor-free control image, generates a pixel-level uncertainty map based on segmentation features using Monte Carlo discarding, and then performs uncertainty-weighted estimation. Intersection operations ensure tumor core purity, and union operations combined with Euclidean distance weights expand the coverage area. The pixel-level uncertainty map reduces the weight of dual-source evidence, and threshold filtering generates a mask to ignore ineffective evidence and high-uncertainty pixels. This transforms multi-source evidence into a continuous prior probability map, providing precise and complete localization guidance for the segmentation decoder. This achieves efficient fusion of weakly supervised evidence and effective suppression of unreliable regions, improving the localization accuracy and boundary integrity of the segmentation task.

[0006] The technical solution adopted by this invention is as follows: This invention provides a combined brain tumor MRI classification method based on domain alignment evidence, the method comprising the following steps:

[0007] Step S1: Predefined, specifically, the MRI image dataset with sufficient annotations is abbreviated as the source domain, and the MRI image dataset with sparse annotations and no annotations is abbreviated as the target domain;

[0008] Step S2: Integration architecture, specifically, a dual-view encoder is constructed using a lightweight ViT backbone network. The dual-view encoder consists of a segment encoder and a classification encoder. A lightweight U-Net network is used to integrate the segment decoder, forming an asymmetric structure of dual encoder and single decoder. A bidirectional cross-attention mechanism is used to establish a feature interaction bridge. The feature interaction bridge is responsible for multi-scale bidirectional information transmission between the segment encoder and the classification encoder. The segment encoder outputs features related to structural geometric segmentation, and the classification encoder outputs features related to global discriminative classification. This makes the features related to structural geometric segmentation output by the segment encoder and the features related to global discriminative classification output by the classification encoder complementary, ultimately generating a shared embedding.

[0009] Step S3: Domain alignment, used to map the shared embeddings to segmentation features and classification features. Specifically, the statistical attributes include mean, variance and covariance. By matching the statistical attributes of the source domain and the target domain, the distributions of the shared embeddings corresponding to the source domain and the target domain are forced to converge.

[0010] Based on statistical properties, adversarial alignment operations are used to obtain segmentation features by making the distribution of geometric segmentation-related features of representation structure in the source domain converge with that in the target domain, and to obtain classification features by making the distribution of global discriminative classification-related features of representation structure in the source domain converge with that in the target domain.

[0011] Step S4: Evidence-guided fusion to obtain heatmaps, differential priors, and pixel-level uncertainty maps;

[0012] Step S5: Uncertainty-weighted estimation based on fusion. Specifically, the differential prior and heatmap are denoted as multi-source weak supervision evidence. First, the multi-source weak supervision evidence is weighted using a pixel-level uncertainty map to suppress interference from low-confidence regions. Then, the multi-source weak supervision evidence is transformed into a continuous prior probability map. The continuous prior probability map not only clarifies the core location of the tumor, but also delineates the fuzzy boundary of the tumor through optimized weights, which is used to adapt to the precise positioning requirements of the segmentation decoder.

[0013] Step S6: Uncertainty image weight reduction, specifically, setting a threshold, filtering continuous prior probability maps to obtain an ignoring mask, the ignoring mask is used to filter pixels that have neither structural difference evidence nor discriminative evidence, and pixels with high inherent uncertainty, and weight reduction processing is performed on unreliable regions of the segmentation features and the classification features according to the ignoring mask;

[0014] Step S7: The segmentation decoder receives segmentation features and classification features while incorporating a continuous prior probability map, performs a decoding operation, and finally completes pixel-level tumor sub-region segmentation, and outputs the uncertainty estimate corresponding to the tumor sub-region.

[0015] Furthermore, in step S4, the evidence-guided fusion specifically includes the following steps:

[0016] Step S41: Heatmap generation, specifically, calculating the weighted gradient of the classification features after domain alignment to the high-level convolutional feature map in the classification encoder to obtain a heatmap. The high-response region in the heatmap is a key region strongly correlated with the category judgment. The heatmap is normalized and used as the initial localization evidence for tumor segmentation under weak supervision to achieve visualized decision-making.

[0017] Step S42: Generate a difference prior to reflect the structural differences between tumor and tumor-free images, transforming the initial localization evidence into a learnable prior. Specifically, a tumor-free control MRI image is reconstructed based on the original MRI image with tumor. The difference between the original MRI image with tumor and the control MRI image is used as the difference prior, providing a boundary-sensitive structural prior for the segmentation task without pixel-level annotation.

[0018] Step S43: Obtain a pixel-level uncertainty map. Specifically, based on the segmentation features, while retaining the Dropout layer in the last two convolutional layers of the segmentation decoder, a Monte Carlo dropout method is used to continuously perform random forward propagation. The variance of the tumor probability output by each pixel during the random forward propagation is calculated. After normalizing the variances of all pixels, they are arranged according to the resolution of the original MRI image to obtain a pixel-level uncertainty map. The pixel-level uncertainty map marks the regions with high uncertainty, i.e., low confidence regions.

[0019] Furthermore, in step S5, the uncertainty-weighted estimation based on fusion specifically includes the following steps:

[0020] Step S51: Intersection, specifically, taking the intersection of the high-response area of ​​the heatmap and the high-response area of ​​the differential prior to ensure the purity of the tumor core location;

[0021] Step S52: Take the union, specifically, integrate the unions of the high response regions of the heatmap and the high response regions of the differential priors that do not overlap, calculate the Euclidean distance from each pixel in the union to the intersection, and use the Euclidean distance as the weight to filter and obtain a continuous prior probability map, ensuring the integrity of the continuous prior probability map.

[0022] This invention provides a combined MRI classification method for brain tumors based on domain alignment evidence. The beneficial effects achieved by this invention using the above scheme are as follows:

[0023] (1) To address the technical problems in brain tumor MRI segmentation, such as domain shift in cross-device data and lack of pixel-level annotation in weakly supervised scenarios, which leads to poor model generalization ability, blurred segmentation boundaries, and easy omission of small lesions, this solution uses a lightweight ViT backbone network to construct a dual-view encoder. Through bidirectional cross-attention, a feature interaction bridge is established to achieve multi-scale bidirectional information complementarity and generate shared embeddings. After domain alignment, the statistical attributes of the source domain and the target domain are matched. By performing adversarial alignment, the feature distribution is forced to converge. Then, through evidence-guided fusion, heatmaps, differential priors, and pixel-level uncertainty maps are generated. After uncertainty weighted estimation, multi-source weakly supervised evidence is transformed into a continuous prior probability map. Combined with ignoring mask filtering of unreliable pixels, the lightweight U-Net segmentation decoder fuses features and prior maps to perform segmentation, achieving a dual improvement in cross-domain robustness and weakly supervised segmentation accuracy, effectively characterizing the tumor core and blurred boundaries, and reducing missed detections and false positives.

[0024] (2) To address the technical problems of insufficient effectiveness of prior guidance caused by the lack of specificity in the fusion of multi-source weakly supervised evidence and the large interference of unreliable regions in traditional segmentation models, this solution uses an evidence-guided fusion process. Based on classification features, a heatmap is generated as the initial localization evidence. The difference prior is obtained by the difference between the original tumor MRI image and the reconstructed tumor-free control image. Based on the segmentation features, a pixel-level uncertainty map is generated using the Monte Carlo discarding method. After uncertainty weighting estimation, the tumor core purity is ensured by the intersection operation, and the coverage is expanded by the union operation combined with Euclidean distance weight. The pixel-level uncertainty map is used to reduce the weight of dual-source evidence. Threshold screening is used to generate a mask to filter out pixels with no effective evidence and high uncertainty. Multi-source evidence is transformed into a continuous prior probability map, which provides accurate and complete localization guidance for the segmentation decoder. This achieves efficient fusion of weakly supervised evidence and effective suppression of unreliable regions, improving the localization accuracy and boundary integrity of the segmentation task. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating a combined brain tumor MRI classification method based on domain alignment evidence provided by the present invention.

[0026] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0028] Example 1, see Figure 1 This invention provides a combined MRI classification method for brain tumors based on domain alignment evidence, the method comprising the following steps:

[0029] Step S1: Predefined, specifically, the MRI image dataset with sufficient annotations is abbreviated as the source domain, and the MRI image dataset with sparse annotations and no annotations is abbreviated as the target domain;

[0030] Step S2: Integration architecture, specifically, a dual-view encoder is constructed using a lightweight ViT backbone network. The dual-view encoder consists of a segment encoder and a classification encoder. A lightweight U-Net network is used to integrate the segment decoder, forming an asymmetric structure of dual encoder and single decoder. A bidirectional cross-attention mechanism is used to establish a feature interaction bridge. The feature interaction bridge is responsible for multi-scale bidirectional information transmission between the segment encoder and the classification encoder. The segment encoder outputs features related to structural geometric segmentation, and the classification encoder outputs features related to global discriminative classification. This makes the features related to structural geometric segmentation output by the segment encoder and the features related to global discriminative classification output by the classification encoder complementary, ultimately generating a shared embedding.

[0031] Step S3: Domain alignment, used to map the shared embeddings to segmentation features and classification features. Specifically, the statistical attributes include mean, variance and covariance. By matching the statistical attributes of the source domain and the target domain, the distributions of the shared embeddings corresponding to the source domain and the target domain are forced to converge.

[0032] Based on statistical properties, adversarial alignment operations are used to obtain segmentation features by making the distribution of geometric segmentation-related features of representation structure in the source domain converge with that in the target domain, and to obtain classification features by making the distribution of global discriminative classification-related features of representation structure in the source domain converge with that in the target domain.

[0033] Step S4: Evidence-guided fusion to obtain heatmaps, differential priors, and pixel-level uncertainty maps;

[0034] Step S5: Uncertainty-weighted estimation based on fusion. Specifically, the differential prior and heatmap are denoted as multi-source weak supervision evidence. First, the multi-source weak supervision evidence is weighted using a pixel-level uncertainty map to suppress interference from low-confidence regions. Then, the multi-source weak supervision evidence is transformed into a continuous prior probability map. The continuous prior probability map not only clarifies the core location of the tumor, but also delineates the fuzzy boundary of the tumor through optimized weights, which is used to adapt to the precise positioning requirements of the segmentation decoder.

[0035] Step S6: Uncertainty image weight reduction, specifically, setting a threshold, filtering continuous prior probability maps to obtain an ignoring mask, the ignoring mask is used to filter pixels that have neither structural difference evidence nor discriminative evidence, and pixels with high inherent uncertainty, and weight reduction processing is performed on unreliable regions of the segmentation features and the classification features according to the ignoring mask;

[0036] Step S7: The segmentation decoder receives segmentation features and classification features while incorporating a continuous prior probability map, performs a decoding operation, and finally completes pixel-level tumor sub-region segmentation, and outputs the uncertainty estimate corresponding to the tumor sub-region.

[0037] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S4, the evidence-guided fusion specifically includes the following steps:

[0038] Step S41: Heatmap generation, specifically, calculating the weighted gradient of the classification features after domain alignment to the high-level convolutional feature map in the classification encoder to obtain a heatmap. The high-response region in the heatmap is a key region strongly correlated with the category judgment. The heatmap is normalized and used as the initial localization evidence for tumor segmentation under weak supervision to achieve visualized decision-making.

[0039] Step S42: Generate a difference prior to reflect the structural differences between tumor and tumor-free images, transforming the initial localization evidence into a learnable prior. Specifically, a tumor-free control MRI image is reconstructed based on the original MRI image with tumor. The difference between the original MRI image with tumor and the control MRI image is used as the difference prior, providing a boundary-sensitive structural prior for the segmentation task without pixel-level annotation.

[0040] Step S43: Obtain a pixel-level uncertainty map. Specifically, based on the segmentation features, while retaining the Dropout layer in the last two convolutional layers of the segmentation decoder, a Monte Carlo dropout method is used to continuously perform random forward propagation. The variance of the tumor probability output by each pixel during the random forward propagation is calculated. After normalizing the variances of all pixels, they are arranged according to the resolution of the original MRI image to obtain a pixel-level uncertainty map. The pixel-level uncertainty map marks the regions with high uncertainty, i.e., low confidence regions.

[0041] Example 3, see Figure 1 This embodiment is based on the above embodiment. In step S5, the uncertainty-weighted estimation based on fusion specifically includes the following steps:

[0042] Step S51: Intersection, specifically, taking the intersection of the high-response area of ​​the heatmap and the high-response area of ​​the differential prior to ensure the purity of the tumor core location;

[0043] Step S52: Take the union, specifically, integrate the unions of the high response regions of the heatmap and the high response regions of the differential priors that do not overlap, calculate the Euclidean distance from each pixel in the union to the intersection, and use the Euclidean distance as the weight to filter and obtain a continuous prior probability map, ensuring the integrity of the continuous prior probability map.

[0044] Example 4, see Figure 1 This embodiment is based on the above embodiment. In step S43, obtaining the pixel-level uncertainty map specifically refers to the following: This embodiment takes a 3.0T MRI brain tumor labeled dataset as the source domain and a 1.5T MRI brain tumor unlabeled dataset as the target domain. First, the segmentation features output by the segmentation encoder are obtained. These features contain the structural geometric information of the brain tumor. Dropout layers are retained in the last two convolutional layers of the segmentation decoder. The Monte Carlo dropout method is used to perform 20 consecutive random forward propagations on an MRI image containing a low-grade glioma in the target domain. After each propagation, the probability (range 0-1) of each pixel in the output image belonging to the tumor region is calculated. The probability of pixels in the tumor core region is concentrated between 0.9 and 0.98, while the probability of pixels at the blurred boundary fluctuates between 0.3 and 0.98. Between -0.7, the pixel probability of normal brain tissue is less than 0.1. The variance of the output probability of each pixel 20 times is calculated. The variance of the tumor core pixel is 0.02-0.05, the variance of the blurred boundary pixel is 0.6-0.8, and the variance of the normal brain tissue pixel is less than 0.03. The variance of all pixels is normalized to the range of 0-1 and rearranged according to the resolution of the original MRI image to generate a pixel-level uncertainty map. In the map, the high variance area corresponds to the blurred boundary of the tumor and the low confidence area, and the low variance area corresponds to the tumor core and normal brain tissue, which provides a precise localization basis for subsequent suppression of interference from unreliable areas.

[0045] Example 5, see Figure 1This embodiment is based on the above embodiment. In steps S51-S52, the uncertainty-weighted estimation based on fusion specifically refers to the following: Taking a 1.5T MRI image containing a low-grade glioma in the target domain as an example, the high response value of the tumor region in the heatmap is concentrated between 0.7 and 1.0, corresponding to the approximate location of the tumor. In the prior difference, the high response area of ​​the structural difference between the tumor and the control image without tumor is 0.6-0.9, accurately reflecting the tumor boundary features. In the pixel-level uncertainty image, the uncertainty value of the tumor core region is less than 0.3, and the uncertainty value of the blurred boundary region is 0.6-0.8. In the intersection operation of step S51, the intersection of the heatmap response value ≥0.7 and the prior difference response value ≥0.6 is extracted. The set region retains only tumor cores confirmed by two types of evidence simultaneously, eliminating suspicious regions supported by a single type of evidence, ensuring that the purity of the tumor core reaches over 95%. In step S52, the union operation integrates the non-overlapping parts of the heatmap and the high-response region of the difference prior, calculates and sets the Euclidean distance from each pixel to the intersection region, sets the weight of distances ≤3 pixels to 0.9, distances 3-5 pixels to 0.5, and distances greater than 5 pixels to 0.1. After reducing the weight of low-confidence regions by combining the pixel-level uncertainty map, a continuous prior probability map is finally generated, which not only clarifies the location of the tumor core, but also completely delineates the fuzzy boundary through gradient weights, providing accurate guidance for the subsequent segmentation decoder.

[0046] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0047] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0048] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A combined MRI classification method for brain tumors based on domain alignment evidence, characterized in that: The method includes the following steps: Step S1: Predefined, specifically, the MRI image dataset with sufficient annotations is abbreviated as the source domain, and the MRI image dataset with sparse annotations and no annotations is abbreviated as the target domain; Step S2: Integration architecture, specifically, a dual-view encoder is constructed using a lightweight ViT backbone network. The dual-view encoder consists of a segment encoder and a classification encoder. A lightweight U-Net network is used to integrate the segment decoder. A bidirectional cross-attention mechanism is used to establish a feature interaction bridge. The feature interaction bridge is responsible for multi-scale bidirectional information transmission between the segment encoder and the classification encoder. The segment encoder outputs features related to structural geometric segmentation, and the classification encoder outputs features related to global discriminative classification. This makes the features related to structural geometric segmentation output by the segment encoder and the features related to global discriminative classification output by the classification encoder complementary, ultimately generating a shared embedding. Step S3: Domain alignment, used to map the shared embeddings to segmentation features and classification features. Specifically, the statistical attributes include mean, variance and covariance. By matching the statistical attributes of the source domain and the target domain, the distributions of the shared embeddings corresponding to the source domain and the target domain are forced to converge. Based on statistical properties, adversarial alignment operations are used to obtain segmentation features by making the distribution of geometric segmentation-related features of representation structure in the source domain converge with that in the target domain, and to obtain classification features by making the distribution of global discriminative classification-related features of representation structure in the source domain converge with that in the target domain. Step S4: Evidence-guided fusion to obtain heatmaps, differential priors, and pixel-level uncertainty maps; Step S5: Uncertainty weighted estimation based on fusion. Specifically, the differential prior and heatmap are denoted as multi-source weak supervision evidence. First, the multi-source weak supervision evidence is weighted using a pixel-level uncertainty map. Then, the multi-source weak supervision evidence is transformed into a continuous prior probability map. Step S6: Uncertainty image weight reduction, specifically, setting a threshold, filtering continuous prior probability maps to obtain an ignore mask, and performing weight reduction processing on unreliable regions of the segmentation features and the classification features based on the ignore mask; Step S7: The segmentation decoder receives segmentation features and classification features while incorporating a continuous prior probability map, performs a decoding operation, and finally completes pixel-level tumor sub-region segmentation, and outputs the uncertainty estimate corresponding to the tumor sub-region.

2. The method for joint brain tumor MRI classification based on domain alignment evidence according to claim 1, characterized in that: In step S4, the evidence-guided fusion specifically includes the following steps: Step S41: Heatmap generation, specifically, calculating the weighted gradient of the classification features after domain alignment to the high-level convolutional feature map in the classification encoder to obtain a heatmap. The high-response region in the heatmap is a key region strongly correlated with the category judgment. The heatmap is normalized and used as the initial localization evidence for tumor segmentation under weak supervision. Step S42: Generate a difference prior to reflect the structural differences between tumor and non-tumor images, and transform the initial localization evidence into a learnable prior. Specifically, reconstruct a control MRI image without tumor based on the original MRI image with tumor, and use the difference between the original MRI image with tumor and the control MRI image as the difference prior. Step S43: Obtain the pixel-level uncertainty map. Specifically, based on the segmentation features, while retaining the Dropout layer in the last two convolutional layers of the segmentation decoder, the Monte Carlo dropout method is used to continuously perform random forward propagation. The variance of the tumor probability output by each pixel during the random forward propagation is calculated. After normalizing the variance of all pixels, they are arranged according to the resolution of the original MRI image to obtain the pixel-level uncertainty map.

3. The method for joint brain tumor MRI classification based on domain alignment evidence according to claim 2, characterized in that: In step S5, the fusion-based uncertainty-weighted estimation specifically includes the following steps: Step S51: Take the intersection, specifically, take the intersection of the high response region of the heatmap and the high response region of the differential prior; Step S52: Take the union, specifically, integrate the unions of the high response regions of the heatmap and the high response regions of the differential priors that do not overlap, calculate the Euclidean distance from each pixel in the union to the intersection, and filter to obtain a continuous prior probability map using the Euclidean distance as the weight.

Images