A bidirectional position-guided medical image segmentation and classification method

Abstract

The application relates to a bidirectional position-guided medical image segmentation and classification method, a multi-task neural network model with bidirectional position guidance is constructed, a forward position guidance map is generated through a classification branch, the classification branch is focused on a key region directly related to image-level discrimination, the forward position guidance map is transmitted to a segmentation branch, the attention ability of the segmentation branch to a lesion region is enhanced, the influence of background interference on a segmentation result is reduced, forward position guidance of classification on segmentation is realized, a pixel-level uncertainty estimation mechanism is introduced into the segmentation branch, the reliability of a segmentation result is screened, a backward position guidance map is constructed in combination with pixel-level uncertainty, the region information fed back to the classification branch is more stable and reliable, the interference of a boundary fuzzy region, a local misclassified region and noise response on classification decision is effectively weakened, backward position guidance of segmentation on classification is realized, more sufficient high-level semantic interaction between the two branches is realized, and the classification accuracy and segmentation consistency are improved.

Description

Technical Field

[0001] This invention relates to the field of medical image classification and segmentation technology, and more specifically, to a bidirectional position-guided method for medical image segmentation and classification. Background Technology

[0002] Medical image classification tasks determine the pathological category of an image, while segmentation tasks locate and delineate lesion regions. These two tasks are naturally related in clinical diagnosis: classification decisions typically rely on the morphology, boundaries, texture, and internal structure of the lesion region, while segmentation also requires lesion saliency and discriminative cues to locate the region. Therefore, joint classification and segmentation methods are gradually becoming an important direction in intelligent medical image analysis.

[0003] Existing joint learning methods mainly fall into two categories: parameter sharing and task interaction. While parameter sharing methods can improve feature representation capabilities through multi-task supervision, they do not adequately utilize high-level semantic interactions between tasks. Although task interaction methods can establish information transmission pathways between classification and segmentation, they do not pay enough attention to the organization and effectiveness of the interaction information itself, resulting in coarse high-level responses with relatively limited stability, thus affecting the full realization of cross-task guidance. Summary of the Invention

[0004] The technical problem this invention aims to solve is how to fully utilize the high-level semantic interaction between segmentation tasks and branching tasks to improve classification accuracy and segmentation consistency.

[0005] This invention provides a bidirectional position-guided medical image segmentation and classification method, comprising the following steps: S1. Acquire medical images, and after annotation and preprocessing, obtain an image dataset, which is then divided into a training set and a validation set. S2 constructs a bidirectional, position-guided multi-task neural network model based on U-Net, which includes a shared encoder and classification and segmentation branches connected to the shared encoder, respectively; wherein... Shared encoders are used to extract shared feature maps from medical images; The classification branch is used to extract classification feature maps from shared feature maps and generate forward position guidance maps through convolutional mapping, which are then passed to the segmentation branch. The classification branch also compresses the classification feature maps based on the forward position guidance maps, generates latent feature maps, and modulates and splices them with the backward position guidance maps passed from the segmentation branch to output the classification results. The segmentation branch is used to extract segmentation feature maps from the shared feature maps, modulate and stitch the segmentation feature maps using the forward position guidance map to output the segmentation result, and then calculate the pixel-level uncertainty map based on the segmentation result and combine it with the segmentation result to obtain the backward position guidance map, which is then passed to the classification branch. S3 constructs a joint loss function with classification accuracy, segmentation consistency and spatial response reliability as optimization objectives, and uses the training set to train the multi-task neural network model; S4 uses a validation set to evaluate the trained multi-task neural network model. After evaluation, the multi-task neural network model is used to analyze the medical images to be detected, and the segmentation and classification results are output.

[0006] Compared with existing technologies, this application has the following advantages: By constructing a multi-task neural network model with bidirectional position guidance, a forward position guidance map is generated through convolutional mapping of the classification branch, enabling the classification branch to focus on key regions directly related to image-level discrimination, and the forward position guidance map is passed to the segmentation branch, thereby enhancing the segmentation branch's ability to focus on lesion regions and reducing the impact of background interference on the segmentation results, thus realizing forward position guidance of classification on segmentation; Next, a pixel-level uncertainty estimation mechanism is introduced into the segmentation branch to screen the reliability of the segmentation results, and a backward position guidance map is constructed in combination with pixel-level uncertainty, making the regional information fed back to the classification branch more stable and reliable, thereby effectively reducing the interference of blurred boundary regions, local misclassified regions and noise response on classification decisions, thus realizing backward position guidance of segmentation on classification. Forward and backward position guidance enable more sufficient high-level semantic interaction between the classification branch and the segmentation branch, improving classification accuracy and segmentation consistency.

[0007] In one possible implementation, the classification branches include sequentially connected: The classification feature extraction layer is used to extract features from the shared feature map to obtain the classification feature map; Bottleneck layer, used to generate forward position guidance map and latent feature map; The backward modulation layer is used to modulate and splice the backward position guidance map and the latent feature map; The classification output header is used to output the classification results.

[0008] In one possible implementation, the bottleneck layer generates a forward position guidance map and a latent feature map through the following operations: Convolutional mapping of the classification feature map yields a forward position guidance map of the same size as the classification feature map; The generated forward position guidance map is passed to the forward modulation layer, and the classification feature map is compressed using the forward position guidance map to obtain the latent feature map: ; In the formula, Represents the latent feature map, This represents a forward positioning guide diagram. Represents the classification feature map, This represents element-wise multiplication. Representation and classification feature maps Gaussian noise tensors with uniform size, mean, and variance.

[0009] Compared with existing technologies, the forward position guidance map enables the multi-task neural network model to automatically learn the importance weight of each spatial location for classification decisions. The forward position guidance map is also represented as a gating response tensor, where high response regions correspond to key lesions and low response regions are suppressed. The forward position guidance map is used to compress the classification feature map, so that high importance regions retain their original features and low importance regions are replaced by noise, thereby reducing redundant information and enhancing the discriminative power of classification features.

[0010] In one possible implementation, the backward modulation layer modulates and splices the backward position guide map and the latent feature map through the following operation: Channel compression is performed on the latent feature map to obtain the high-level classification feature map; The backward position guidance map is adjusted to the same spatial size as the high-level classification feature map; then the adjusted backward position guidance map is used to modulate the high-level classification feature map position by position; then the high-level classification feature map is spliced ​​together to obtain the backward-guided classification feature map. The classification output head uses a fully connected layer to process the backward-guided classification feature map and output the classification result.

[0011] Compared with existing technologies, this method utilizes the adjusted backward position guidance map to perform position-by-position modulation on the high-level classification feature map, achieving spatial attention guidance of the segmentation branch to the classification branch. Regions considered stable and reliable by the segmentation branch are enhanced, while uncertain regions are suppressed. Channel splicing with the high-level classification feature map preserves the global discriminative information learned by the classification branch itself and incorporates the locally reliable priors provided by the segmentation branch. The two complement each other, improving the representational ability of the classification features. Finally, the classification result is output through the fully connected layer of the classification output head, making the classification decision more dependent on stable and consistent feature responses within the lesion region. This reduces the interference of boundary blurring, artifacts, or noise on the classification, thereby improving the robustness and accuracy of the classification.

[0012] In one possible implementation, the split branches comprise sequentially connected: The segmentation feature extraction layer is used to extract features from the shared feature map to obtain the segmentation feature map; The forward modulation layer, connected to the bottleneck layer, is used to modulate and splice the forward position guidance map and the segmentation feature map; The segmentation output header is used to output the segmentation results; An estimation layer is used to calculate a pixel-level uncertainty map based on the segmentation results; It also includes a guidance layer connected to the segmentation output head and the estimation layer. The guidance layer is used to generate a backward position guidance map by combining the segmentation results with the pixel-level uncertainty map. The guidance layer is connected to the backward modulation layer. In one possible implementation, the forward modulation layer modulates and splices the forward position guidance map and the segmentation feature map through the following operation: Upsample the forward position guidance map to the same spatial size as the segmentation feature map; After the segmentation feature map is modulated position by position using the forward position guidance map, it is then concatenated with the segmentation feature map to obtain the concatenated segmentation feature map; The segmentation output head outputs the segmentation result through the following operations: Channel compression is performed on the spliced ​​segmentation feature map to obtain the segmentation feature map after forward position guidance; Based on the segmentation feature map guided by the forward position, the segmentation result is output by the decoder.

[0013] Compared with existing technologies, the segmentation branch achieves spatial attention guidance from the classification branch to the segmentation branch by upsampling the forward location guidance map and multiplying it element-wise with the segmentation feature map. That is, regions that the classification branch considers important are enhanced, while unimportant regions are suppressed, thereby improving the localization accuracy of the segmentation branch for key lesion regions. Then, channel concatenation is performed with the segmentation feature map, which not only preserves the original expressive power of the segmentation branch but also introduces classification priors, resulting in stronger information complementarity. Finally, the segmentation result is output through the decoder, ensuring that the segmentation result restores spatial details and is suitable for fine lesion depiction in medical images.

[0014] In one possible implementation, the estimation layer calculates a pixel-level uncertainty map based on the segmentation results by: Based on segmentation results Define Dirichlet parameters : ; in, Indicates pixel position, Indicates the segmentation category; Based on Dirichlet parameters Calculate pixel position Total strength of evidence : ; in, Indicates the total number of segmentation categories; Based on pixel position Total strength of evidence Calculate the position of this pixel pixel-level uncertainty : ; Build a pixel-level uncertainty map from the pixel-level uncertainties at all pixel locations. ; The guiding layer uses an exponential decay function to convert pixel-level uncertainty into reliability weights, and then performs element-wise weighting with the segmentation results to obtain the backward position guidance map. : ; in, This indicates the segmentation result.

[0015] Compared to existing technologies, estimating pixel-level uncertainty using evidence theory (Dirichlet distribution) quantifies the credibility of each pixel's prediction: stronger evidence (more certain segmentation) results in lower uncertainty; weaker evidence (fuzzy boundaries or misclassification) results in higher uncertainty; and an exponential decay function is employed. Uncertainty is converted into reliability weights, giving higher weights to high-reliability regions (low uncertainty) and weights to low-reliability regions that approach zero; the resulting backward positioning guidance map. It also includes the shape information and reliability confidence of the segmentation results, representing a stable and clean lesion region prior; the backward location guidance map The data is passed to the classification branch, effectively shielding unreliable, ambiguous, or erroneous regional feedback in the segmentation branch, thus avoiding misleading classification decisions.

[0016] In one possible implementation, the joint loss function constructed in step S3 is: ; In the formula, This represents the classification loss of the classification branch. This represents the partition supervision loss for the partitioned branches. This represents the modeling loss used to constrain the segmentation results and their pixel-level uncertainties. , and All are weighting coefficients.

[0017] Compared with existing technologies, the segmentation supervision loss directly constrains the consistency between the final segmentation result and the true label, ensuring segmentation accuracy. The modeling loss enables the multi-task neural network model to actively estimate the confidence of its own predictions while learning segmentation, providing a reliable basis for the backward guidance graph.

[0018] In one possible implementation, the classification loss Including classification cross-entropy loss and bottleneck constraint terms: ; In the formula, This indicates the output classification result. This indicates the actual category label being marked. For balance coefficient, Represents the classification cross-entropy loss: , Indicates the total number of categories; Bottleneck constraint terms: , For latent feature maps, For classification feature maps, Represented by classification feature map Induced latent feature map Conditional distribution, Representing latent feature maps The variational approximation distribution, This represents the Kullback–Leibler divergence.

[0019] In one possible implementation, the segmentation supervision loss Cross-entropy loss is used: ; In the formula, This indicates the actual segmentation label. This indicates the splitting result output by the splitting branch.

[0020] In one possible implementation, modeling loss Used to jointly constrain the segmentation results and the estimated pixel-level uncertainties: ; In the formula, This represents a segmentation supervision term based on evidence-based learning. , Represents the digamma function. Indicates the pixel position And the segmentation category is The actual category label is marked there. This represents the regularization constraint term on the pixel-level evidential parameter distribution. This represents the consistency constraint with the actual segmentation label. and These are the weighting coefficients. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the application process. Figure 2 This is a schematic diagram of the structure of the bidirectional position-guided multi-task neural network model of this application. Detailed Implementation

[0022] First, those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the embodiments of this application and are not intended to limit the scope of protection of the embodiments of this application. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0023] In the description of the embodiments of this application, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of this application based on the specific circumstances.

[0024] In the embodiments of this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0025] The present application will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0026] See Figures 1-2 As shown in the illustration, this application discloses a bidirectional position-guided medical image segmentation and classification method. This method constructs a bidirectional guided multi-task neural network model based on U-Net and performs end-to-end optimization using a joint loss function. This application uses the lung nodule segmentation and classification task in lung CT images as an example for illustration, specifically including the following steps: S1. Acquire medical images, and after annotation and preprocessing, obtain an image dataset which is then divided into a training set and a validation set. In this embodiment, a dataset of lung CT images is obtained. Each image is labeled with pixel-level lesion regions (segmentation labels) and image-level category labels (classification labels, such as "nodule positive / negative") by a professional physician. The images are preprocessed, including normalization, scaling to a uniform size (e.g., 256×256), and data augmentation (random rotation, flipping, etc.). The preprocessed dataset is divided into a training set and a validation set in an 8:2 ratio.

[0027] S2, a bidirectional position-guided multi-task neural network model is constructed based on U-Net, which includes a shared encoder and classification and segmentation branches respectively connected to the shared encoder; wherein: A shared encoder is used to extract shared feature maps from medical images. The shared encoder employs the encoding portion of U-Net and consists of four downsampling blocks. Each downsampling block contains two consecutive 3×3 convolutional layers (each followed by BatchNorm and ReLU) and a 2×2 max-pooling layer. The input lung CT image (size 256×256×1) is processed by the shared encoder to output a shared feature map. Its spatial dimensions are 16×16, and the number of channels is 512.

[0028] The classification branch is used to extract classification feature maps from the shared feature maps and generate a forward position guide map through convolutional mapping, which is then passed to the segmentation branch. The classification branch also compresses the classification feature maps based on the forward position guide map, generates a latent feature map, and then modulates and splices it with the backward position guide map passed from the segmentation branch to output the classification result.

[0029] The classification branch consists of a classification feature extraction layer, a bottleneck layer, a backward modulation layer, and a classification output head, connected in sequence; where: The classification feature extraction layer is used to extract features from the shared feature map to obtain the classification feature map. The classification feature extraction layer consists of three consecutive 3×3 convolutional layers (each followed by BatchNorm and ReLU); feature extraction is performed on the shared feature map X to obtain the classification feature map. The dimensions are 16×16×256.

[0030] The bottleneck layer first processes the classification feature map Convolutional mapping is performed to obtain the classification feature map. Forward position guidance map of the same size Specifically, the bottleneck layer first passes through a 1×1 convolutional mapping layer. and Sigmoid activation function Generate and classify feature maps Forward position guidance map of the same size , Next, the bottleneck layer generates a forward position guidance map. While passing the data to the split branch, utilize the forward position guidance graph. Classification feature map Information bottleneck compression is performed to obtain latent feature maps. ;in, This represents element-wise multiplication. Indicates and Gaussian noise tensors of uniform size with the same mean and variance; this latent feature map Simultaneously serving as a forward positioning guide map The carrier and the basis for backward positioning guidance.

[0031] The backward modulation layer first modulates the latent feature map Channel compression is performed (reducing the number of channels from 256 to 128 using a 1×1 convolution) to obtain high-level classification feature maps. ,in, This indicates channel compression; next, the backward position guidance diagram passed from the split branch is used. Adjust to high-level classification feature map For the same space size, it is denoted as Reuse High-level classification feature map Perform position-by-position modulation, then high-level classification feature map Feature concatenation is performed to obtain the classification feature map after back-guided mapping. In the formula, This represents element-wise multiplication. This indicates channel splicing.

[0032] The classification output head will guide the classification feature map. The classification output header uses a fully connected layer to output two probabilities (positive / negative nodules) to obtain the classification result. : ; In the formula, This indicates a fully connected layer.

[0033] The segmentation branch is used to extract segmentation feature maps from the shared feature maps, modulate and stitch the segmentation feature maps using the forward position guidance map to output the segmentation result, and then calculate the pixel-level uncertainty map based on the segmentation result and combine it with the segmentation result to obtain the backward position guidance map, which is then passed to the classification branch.

[0034] The segmentation branch includes a segmentation feature extraction layer, a forward modulation layer, a segmentation output head, and an estimation layer connected in sequence, and also includes a guiding layer connected to the segmentation output head and the estimation layer; the bottleneck layer is connected to the forward modulation layer, and the guiding layer is connected to the backward modulation layer; wherein: The segmentation feature extraction layer is used to extract features from the shared feature map to obtain the segmentation feature map. The segmentation feature extraction layer has the same structure as the classification feature extraction layer, consisting of three consecutive 3×3 convolutional layers, which extract segmentation feature maps from the shared feature map X. .

[0035] The forward modulation layer passes the forward position guidance map from the classification branch. Upsampled and segmented feature maps With the same spatial dimensions, the forward position guidance diagram is used again. For segmentation feature maps After position-by-position modulation and segmentation feature map The spliced ​​segmentation feature map is obtained by splicing. (Number of channels: 512); ;in, This indicates channel splicing.

[0036] The split output head first uses 3 consecutive... Convolutional layers splice segmentation feature maps Channel compression is performed to obtain the segmentation feature map after forward position guidance. ;in, This indicates channel compression; the number of channels is restored to 256. Next, based on the guided segmentation feature map... The segmentation result is output through the decoder. The segmentation output head of this embodiment adopts a U-Net decoder structure, containing four upsampling blocks. Each upsampling block includes a 2×2 transposed convolution and two 3×3 convolutions, which are concatenated with the corresponding layers of the shared encoder through skip connections; the final output segmentation result is... ;in, , Indicates the number of segmentation categories. This indicates the decoder.

[0037] Based on the segmentation results, the estimation layer calculates a pixel-level uncertainty map through the following operations: Based on segmentation results Define Dirichlet parameters ;in, Indicates pixel position, Indicates the segmentation category; Based on Dirichlet parameters Calculate pixel position Total strength of evidence ; Based on pixel position Total strength of evidence Calculate the position of this pixel pixel-level uncertainty ; Build a pixel-level uncertainty map from the pixel-level uncertainties at all pixel locations. .

[0038] The guiding layer uses an exponential decay function to convert pixel-level uncertainty into reliability weights, which are then compared with the segmentation results. Perform element-wise weighted summation to obtain the backward position guidance graph. The backward position guidance diagram It is passed to the backward modulation layer.

[0039] S3 constructs a joint loss function with classification accuracy, segmentation consistency, and spatial response reliability as optimization objectives, and uses the training set to train the multi-task neural network model.

[0040] In this embodiment, to ensure that the classification branch can correctly identify the pathological category of the medical image and the segmentation branch can accurately depict the lesion region, consistent with the real annotation; and by jointly constraining the initial segmentation result and its pixel-level uncertainty, the multi-task neural network model can evaluate the credibility of its own predictions, thereby providing a stable and reliable region prior for the backward location guidance map, improving the robustness and interpretability of the overall model; a joint loss function is constructed by combining the classification loss, the segmentation supervision loss, and the modeling loss of the segmentation result and its pixel-level uncertainty: ; In the formula, This represents the classification loss of the classification branch. This represents the partition supervision loss for the partitioned branches. This represents the modeling loss used to constrain the segmentation results and their pixel-level uncertainties. , and These are all weighting coefficients. In this embodiment, the parameters are adjusted based on the validation set. , , =0.1.

[0041] The classification loss Including classification cross-entropy loss and bottleneck constraint terms: ; In the formula, This indicates the output classification result. This indicates the actual category label being marked. For balance coefficient, Represents the classification cross-entropy loss: , Indicates the total number of categories; Bottleneck constraint terms: , For latent feature maps, For classification feature maps, Represented by classification feature map Induced latent feature map Conditional distribution, Representing latent feature maps The variational approximation distribution is taken as the standard normal distribution. This represents the Kullback–Leibler divergence.

[0042] The segmentation supervision loss Cross-entropy loss is used: ; In the formula, This indicates the actual segmentation label. This indicates the splitting result output by the splitting branch.

[0043] Modeling loss Used to jointly constrain the segmentation results and the estimated pixel-level uncertainties: ; In the formula, This represents a segmentation supervision term based on evidence-based learning. , Represents the digamma function. Indicates the pixel position And the category label is The actual category label is marked there. This represents a regularization constraint term for the pixel-level evidential parameter distribution. This regularization constraint term is existing technology and will not be elaborated upon here. This represents the consistency constraint with the actual segmentation label. and As a weighting coefficient, in the embodiments of this application, , .

[0044] The multi-task neural network described above was trained using the training set. The optimizer was Adam, with an initial learning rate of 1e-4, a batch size of 8, and a training duration of 100 epochs. After each epoch, classification accuracy and the Dice coefficient for segmentation were calculated on the validation set. The model parameters that best performed on the validation set were saved after training.

[0045] After evaluation, the model is deployed for the automatic analysis of clinical lung CT images: input a CT image to be tested, and the model simultaneously outputs whether nodules are present in the image (classification result) and the precise location and outline of the nodules (segmentation result), assisting doctors in diagnosis.

[0046] S4 uses a validation set to evaluate the trained multi-task neural network model. After evaluation, the multi-task neural network model is used to analyze the medical images to be detected, and the segmentation and classification results are output.

[0047] In the description of the embodiments of this application, it should be noted that the terms "inner" and "outer" and other terms indicating direction or positional relationship are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and does not indicate or imply that the device or component must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this application.

[0048] In the description of this application, the references to terms such as "an embodiment," "some embodiments," "in this embodiment," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0049] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A bidirectional position-guided medical image segmentation and classification method, characterized in that, Includes the following steps: S1. Acquire medical images, and after annotation and preprocessing, obtain an image dataset, which is then divided into a training set and a validation set. S2 constructs a bidirectional, position-guided multi-task neural network model based on U-Net, which includes a shared encoder and classification and segmentation branches connected to the shared encoder, respectively; wherein... Shared encoders are used to extract shared feature maps from medical images; The classification branch is used to extract classification feature maps from the shared feature maps and generate a forward position guide map through convolutional mapping, which is then passed to the segmentation branch. The classification branch also compresses the classification feature maps based on the forward position guide map, generates a latent feature map, and then modulates and splices it with the backward position guide map passed from the segmentation branch to output the classification result. The segmentation branch is used to extract segmentation feature maps from the shared feature maps, modulate and stitch the segmentation feature maps using the forward position guidance map to output the segmentation result, and then calculate the pixel-level uncertainty map based on the segmentation result and combine it with the segmentation result to obtain the backward position guidance map, which is then passed to the classification branch. S3: A joint loss function is constructed by combining the classification loss, the segmentation supervision loss, and the modeling loss of the segmentation result and its pixel-level uncertainty. The multi-task neural network model is then trained using the training set. S4 uses a validation set to evaluate the trained multi-task neural network model. After evaluation, the multi-task neural network model is used to analyze the medical images to be detected, and the segmentation and classification results are output.

2. The bidirectional position-guided medical image segmentation and classification method according to claim 1, characterized in that, The classification branches include the following connected sequentially: The classification feature extraction layer is used to extract features from the shared feature map to obtain the classification feature map; Bottleneck layer, used to generate forward position guidance map and latent feature map; The backward modulation layer is used to modulate and splice the backward position guidance map and the latent feature map; The classification output header is used to output the classification results.

3. The bidirectional position-guided medical image segmentation and classification method according to claim 2, characterized in that, The bottleneck layer generates a forward position guidance map and a latent feature map through the following operations: Convolutional mapping of the classification feature map yields a forward position guidance map of the same size as the classification feature map; The generated forward position guidance map is passed to the forward modulation layer, and the classification feature map is compressed using the forward position guidance map to obtain the latent feature map: ； In the formula, Represents the latent feature map, This represents a forward positioning guide diagram. Represents the classification feature map, This represents element-wise multiplication. Representation and classification feature maps Gaussian noise tensors with uniform size, mean, and variance.

4. The bidirectional position-guided medical image segmentation and classification method according to claim 2, characterized in that, The backward modulation layer modulates and splices the backward position guidance map and the latent feature map through the following operations: Channel compression is performed on the latent feature map to obtain the high-level classification feature map; The backward position guidance map is adjusted to the same spatial size as the high-level classification feature map; then the adjusted backward position guidance map is used to modulate the high-level classification feature map position by position. Then, the feature map is concatenated with the high-level classification feature map to obtain the back-guided classification feature map. The classification output head uses a fully connected layer to process the backward-guided classification feature map and output the classification result.

5. The bidirectional position-guided medical image segmentation and classification method according to claim 2, characterized in that, The segmented branches include sequentially connected: The segmentation feature extraction layer is used to extract features from the shared feature map to obtain the segmentation feature map; The forward modulation layer, connected to the bottleneck layer, is used to modulate and splice the forward position guidance map and the segmentation feature map; The segmentation output header is used to output the segmentation results; An estimation layer is used to calculate a pixel-level uncertainty map based on the segmentation results; It also includes a guiding layer connected to the segmentation output head and the estimation layer. The guiding layer is used to generate a backward position guiding map by combining the segmentation results with the pixel-level uncertainty map. The guiding layer is connected to the backward modulation layer.

6. The bidirectional position-guided medical image segmentation and classification method according to claim 5, characterized in that, The forward modulation layer modulates and splices the forward position guidance map and the segmentation feature map through the following operations: Upsample the forward position guidance map to the same spatial size as the segmentation feature map; After the segmentation feature map is modulated position by position using the forward position guidance map, it is then concatenated with the segmentation feature map to obtain the concatenated segmentation feature map; The segmentation output head outputs the segmentation result through the following operations: Channel compression is performed on the spliced ​​segmentation feature map to obtain the segmentation feature map after forward position guidance; Based on the segmentation feature map guided by the forward position, the segmentation result is output by the decoder.

7. The bidirectional position-guided medical image segmentation and classification method according to claim 5, characterized in that, The estimation layer calculates a pixel-level uncertainty map based on the segmentation results through the following operations: Based on segmentation results Define Dirichlet parameters : ； in, Indicates pixel position, Indicates the segmentation category; Based on Dirichlet parameters Calculate pixel position Total strength of evidence : ； in, Indicates the total number of segmentation categories; Based on pixel position Total strength of evidence Calculate the position of this pixel pixel-level uncertainty : ； Build a pixel-level uncertainty map from the pixel-level uncertainties at all pixel locations. ; The guiding layer uses an exponential decay function to convert pixel-level uncertainty into reliability weights, and then performs element-wise weighting with the segmentation results to obtain the backward position guidance map. : ； in, This indicates the segmentation result.

8. The bidirectional position-guided medical image segmentation and classification method according to claim 7, characterized in that, The joint loss function constructed in step S3 is: ； In the formula, This represents the classification loss of the classification branch. This represents the partition supervision loss of the partitioned branch. This represents the modeling loss used to constrain the segmentation results and their pixel-level uncertainties. , and All are weighting coefficients.

9. The bidirectional position-guided medical image segmentation and classification method according to claim 8, characterized in that, The classification loss Including classification cross-entropy loss and bottleneck constraint terms: ； In the formula, This indicates the output classification result. This indicates the actual category label being marked. For balance coefficient, Represents the classification cross-entropy loss: , Indicates the total number of categories; Bottleneck constraint terms: , For latent feature maps, For classification feature maps, Represented by classification feature map Induced latent feature map Conditional distribution, Representing latent feature maps The variational approximation distribution, Indicates the Kullback–Leibler divergence; The segmentation supervision loss Employing multivariate cross-entropy loss: ； In the formula, This indicates the actual segmentation label. This indicates the splitting result output by the splitting branch.

10. The bidirectional position-guided medical image segmentation and classification method according to claim 9, characterized in that, Modeling loss Used to jointly constrain the segmentation results and the estimated pixel-level uncertainties: ； In the formula, This represents a segmentation supervision term based on evidence-based learning. , Represents the digamma function. Indicates the pixel position And the segmentation category is The actual category label is marked there. This represents the regularization constraint term on the pixel-level evidential parameter distribution. This represents the consistency constraint with the actual segmentation label. and These are the weighting coefficients.

Images