A training method, a detection method and a system of a cerebral hemorrhage lesion detection model

Abstract

The application discloses a cerebral hemorrhage lesion detection model training method, a detection method and a system, and belongs to the technical field of medical image detection; a double-path supervision signal of a relative class activation map and a prototype class activation map is constructed; a relative class activation map is generated by subtracting the maximum logical prediction score of a background area from the logical prediction score map of a brain image pixel by pixel, effectively suppressing the prediction score smoothing problem caused by lesion and background diversity. The relative class activation map highlights the relative difference between the foreground and background responses, significantly sharpening the positioning response map, enhancing its robustness to threshold. By forcing the relative class activation map and the prototype class activation map to be consistent, the double calibration and refinement of the model positioning behavior are realized, not only providing self-consistency verification for the model, but also effectively correcting the positioning ambiguity problem caused by smooth activation, and achieving accurate cerebral hemorrhage detection without pixel-level labeling.

Description

A training method, detection method, and system for a brain hemorrhage lesion detection model. Technical Field

[0001] This application belongs to the field of medical image detection technology, and more specifically, relates to a training method for a brain hemorrhage lesion detection model, a brain hemorrhage lesion detection method and system. Background Technology

[0002] Acute intracerebral hemorrhage is a leading cause of death and disability worldwide, posing a serious threat to public health. Treatment of acute intracerebral hemorrhage is strictly time-dependent. Therefore, if stroke can be accurately diagnosed and lesion areas segmented based on initial brain CT scans, doctors can take appropriate clinical measures according to the patient's different risk levels, which will greatly improve the timeliness, effectiveness, and targeted nature of treatment.

[0003] Currently, most existing brain hemorrhage detection algorithms are based on a fully supervised learning paradigm, which assumes that each sample has a corresponding pixel-level lesion label. However, in real-world medical scenarios, constructing a sufficient and high-quality dataset for brain hemorrhage image detection is extremely challenging. Fine-grained annotation of these brain hemorrhage images is not only tedious and time-consuming but also requires extensive expertise. Therefore, developing weakly supervised brain hemorrhage lesion detection methods that can reduce annotation costs is crucial.

[0004] Given that slice-level labels are the cheapest to obtain, researching a method for accurate brain hemorrhage detection using these weak labels is of great significance. A key challenge is effectively transferring slice-level supervision signals to the pixel level. Existing weakly supervised detection methods largely rely on Class Activation Maps (CAMs) and their variants to generate pixel-level pseudo-labels. These methods are based on the assumption that the model's classification decision for the entire slice should be dominated by the lesion region. Therefore, reverse localization can roughly delineate the lesion region. Furthermore, some studies have attempted to optimize CAMs through strategies such as erasure and expansion to obtain a more complete lesion region. However, in brain hemorrhage segmentation tasks, these methods often fail to achieve ideal results. This is because the contrast between stroke lesions and normal brain tissue in plain CT scans is poor, resulting in scattered and blurry responses in the CAM-generated regions, making it difficult to focus on the true lesion boundaries. Simultaneously, brain hemorrhage lesions exhibit high diversity in location, shape, and size. This morphological heterogeneity makes it easy for the model to learn non-critical features, leading to false activations. The combined effect of these challenges results in overly smooth logical prediction scores generated based on slice-level annotations. This smoothing not only makes pseudo-labels extremely sensitive to threshold selection, but also makes them more prone to serious problems such as region omission (undersegmentation) and background misidentification (oversegmentation), making it impossible to accurately detect cerebral hemorrhage. Summary of the Invention

[0005] In view of the above-mentioned defects or improvement needs of the existing technology, this application provides a training method for a brain hemorrhage lesion detection model, a brain hemorrhage lesion detection method and system, the purpose of which is to achieve accurate brain hemorrhage detection without pixel-level annotation.

[0006] To achieve the above objectives, firstly, this application provides a training method for a brain hemorrhage lesion detection model, comprising:

[0007] Brain image samples from the training set are input in batches into the brain hemorrhage lesion detection model for multiple rounds of training. The brain hemorrhage lesion detection model includes a feature extraction module, a diagnosis module, and a localization module. Each round of training includes:

[0008] Each brain image sample T in the current batch is input into the brain hemorrhage lesion detection model. The feature extraction module is used to extract features from the brain image sample T to obtain a feature map F. The diagnosis module is used to map the feature map F to a logistic prediction score map L, thereby obtaining the classification result of whether the brain image is a brain hemorrhage image. When the brain image T is a brain hemorrhage image, the localization module is used to process the logistic prediction score map L through the ReLU function and normalization process in sequence to obtain a class activation map, thereby obtaining the detection result of the brain hemorrhage lesion region in the brain image.

[0009] Brain image samples classified as brain hemorrhage images in the current batch are used as simulated brain hemorrhage images; for each simulated brain hemorrhage image... : its feature map Based on its class activation graph Weighted average pooling is used to obtain the corresponding foreground features. Its feature map Based on its background class activation graph Weighted average pooling is used to obtain the corresponding background features. ; ;calculate Feature map The features at each pixel location are respectively related to and The similarity is used to obtain the corresponding foreground similarity map. Similarity map with background ;Will and After subtraction, the result is processed by the ReLU function and then normalized to obtain... The corresponding prototype class activation graph; for The position of each pixel in the comparison and The similarity value at the corresponding position is used to determine the region type corresponding to the larger similarity value, thus obtaining the region type to which the pixel belongs. The corresponding pseudo-label image ;based on Sure Corresponding Logistic Prediction Score Chart Background area in to obtain The maximum logistic prediction score in; After subtracting the maximum logical prediction score from each pixel, the result is processed sequentially using the ReLU function and then normalized to obtain... The corresponding relative class activation graph;

[0010] The brain hemorrhage lesion detection model is trained by minimizing the classification loss and consistency loss of the current batch; the consistency loss represents the sum of the difference loss between the prototype class activation map and the relative class activation map corresponding to each simulated brain hemorrhage image in the current batch.

[0011] More preferably, while minimizing the classification loss and consistency loss of the current batch, the prototype difference loss of the current batch is also minimized; wherein, the prototype difference loss represents the foreground feature set. The set of foreground and background features in The sum of the similarities of each background feature in the image; This is the set of foreground features of each simulated brain hemorrhage image in the current batch; This is a set of background features of each simulated brain hemorrhage image in the current batch.

[0012] More preferably, the above-mentioned foreground features and background features They are respectively:

[0013]

[0014]

[0015] Where H and W are feature maps respectively. The number of rows and columns; Representation of feature map The pixel in the i-th row and j-th column; Representation of class activation graph The pixel in the i-th row and j-th column; Represents the background class activation graph The pixel in the i-th row and j-th column.

[0016] More preferably, in the feature map Based on background class activation graph Before performing weighted average pooling, it is also necessary to... Perform a random discard operation.

[0017] More preferably, the expression for the above-mentioned prototype difference loss is:

[0018]

[0019] Where N is the number of simulated brain hemorrhage images in the current batch.

[0020] More preferably, the feature extraction module is used to extract feature maps at multiple scales from brain image sample T, scale them to the same preset size, and then stitch them together in the channel dimension to obtain feature map F.

[0021] Secondly, this application provides a method for detecting cerebral hemorrhage lesions, including:

[0022] The brain image to be detected is input into the brain hemorrhage lesion detection model to obtain the brain hemorrhage lesion detection results;

[0023] The cerebral hemorrhage lesion detection model is trained using the training method provided in the first aspect of this application; the cerebral hemorrhage lesion detection results include: the classification result of whether the brain image is a cerebral hemorrhage image, and the detection result of the cerebral hemorrhage lesion region when the brain image is a cerebral hemorrhage image.

[0024] Thirdly, this application provides a brain hemorrhage lesion detection system, comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the method provided in the second aspect of this application.

[0025] Fourthly, this application provides an electronic device, including: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the method provided in the first or second aspect of this application.

[0026] Fifthly, this application also provides a computer-readable storage medium comprising a stored computer program, wherein the computer program, when executed by a processor, controls the device containing the storage medium to perform the method provided in the first or second aspect of this application.

[0027] Sixthly, this application also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the method provided in the first or second aspect of this application.

[0028] In summary, the above-described technical solutions conceived in this application can achieve the following beneficial effects:

[0029] 1. This application provides a training method for a brain hemorrhage lesion detection model, constructing a dual-path supervision signal consisting of a relative activation map and a prototype activation map. The core of this mechanism lies in generating the relative activation map by subtracting the maximum logical prediction score of the background region from the logical prediction score map of the brain image pixel by pixel. This effectively suppresses the prediction score smoothing problem caused by the diversity of lesions and the background. The relative activation map significantly sharpens the localization response map by highlighting the relative difference between the foreground and background responses, enhancing its robustness to thresholds. By forcing the sharpened relative activation map to maintain consistency with the prototype activation map obtained based on prototype feature matching, dual calibration and refinement of the model's localization behavior are achieved. This process not only provides self-consistency verification for the model but also effectively corrects the localization ambiguity problem caused by smooth activation, significantly alleviating region omissions and background misidentification, thereby improving the accuracy of lesion localization boundaries and the integrity of the region. This enables accurate brain hemorrhage detection without pixel-level annotation.

[0030] 2. Furthermore, the training method for the cerebral hemorrhage lesion detection model provided in this application minimizes both classification loss and consistency loss, as well as prototype difference loss. By minimizing the similarity between foreground and background features, it explicitly drives the model to have highly discriminative feature representations, forcing the model to effectively separate foreground (lesion) features from background (normal tissue) features in the feature space. This training mechanism imposes strong constraints at the feature similarity level, guiding the model to focus on the essential features of stroke lesions and suppressing interference from irrelevant background information, thereby further improving the model's ability to identify lesions and enhancing the accuracy of cerebral hemorrhage detection.

[0031] 3. Furthermore, in the training method of the cerebral hemorrhage lesion detection model provided in this application, the feature extraction module is used to extract feature maps of brain image samples at multiple scales, scale them to the same preset size, and then stitch them together in the channel dimension to obtain feature maps; by integrating and fusing the contextual information of brain images at multiple scales through the feature extraction module, a fusion feature map with rich spatial details and high-level semantics is constructed, which further improves the accuracy of the model in detecting cerebral hemorrhage. Attached Figure Description

[0032] Figure 1 is a schematic diagram of the training process of the cerebral hemorrhage lesion detection model provided in Embodiment 1 of this application.

[0033] Figure 2 is a schematic diagram of all relevant output results during the training process provided in Embodiment 1 of this application.

[0034] Figure 3 is a probability distribution diagram of the prediction without using consistency loss during the training process provided in Embodiment 1 of this application.

[0035] Figure 4 is a probability distribution diagram of the prediction using consistency loss during the training process provided in Embodiment 1 of this application. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. Furthermore, the technical features involved in the various embodiments described below can be combined with each other as long as they do not conflict with each other.

[0037] Under weak supervision, the model relies solely on image-level labels for training, lacking direct supervision with pixel-level annotations. This leads to difficulties in accurately distinguishing lesions from normal tissue in complex brain tissue backgrounds. The generated logistic prediction scores often exhibit smoothing issues, causing activation responses to diffuse and blur at lesion boundaries. This results in threshold sensitivity to localization results, easily leading to both undersegmentation (incomplete lesion region identification) and oversegmentation (background misclassified as lesion). To address these smoothing issues, this application provides a training method for a brain hemorrhage lesion detection model. The core idea is to construct a relative class activation map as a key supervision signal, in addition to traditional classification supervision. This map is generated by subtracting the maximum logistic prediction score of the background region pixel by pixel from the logistic prediction score map of the brain image, thereby sharpening the response difference between the foreground and background and enhancing the robustness of the localization response map to threshold selection. Furthermore, consistency loss constraints are used to force the relative class activation map to maintain consistency with the prototype class activation map generated based on prototype matching, achieving dual calibration and refinement of lesion localization. This mechanism fundamentally constrains the distribution smoothness problem in the feature space, forcing the model to learn more discriminative feature representations, significantly improving the clarity of lesion boundaries and regional integrity, and ultimately achieving high-precision lesion localization and classification using only image-level labels.

[0038] The following is an example.

[0039] Example 1

[0040] This embodiment provides a training method for a brain hemorrhage lesion detection model, including:

[0041] Brain image samples from the training set were input into the cerebral hemorrhage lesion detection model in batches for multiple rounds of training;

[0042] The training set includes multiple brain image samples; each image sample carries only a classification label indicating whether it is a brain hemorrhage image, without carrying pixel-level lesion annotation information.

[0043] The brain hemorrhage lesion detection model includes: a feature extraction module, a diagnostic module, and a localization module; each round of training includes the following parts:

[0044] 1) Input each brain image sample T in the current batch into the brain hemorrhage lesion detection model; wherein, the feature extraction module is used to extract features from the brain image sample T to obtain feature map F; the diagnosis module is used to map feature map F to logistic prediction score map L (the pixel value at each pixel position in the logistic prediction score map is the logistic prediction score at the corresponding pixel position in the input feature map F that is predicted to belong to the foreground region (i.e., the brain hemorrhage region), thereby obtaining the classification result of whether the brain image is a brain hemorrhage image; the localization module is used to process the logistic prediction score map L through the ReLU function and normalization process in sequence when the brain image T is a brain hemorrhage image to obtain the class activation map, thereby obtaining the detection result of the brain hemorrhage lesion region in the brain image;

[0045] It should be noted that the feature extraction module can use CNN models, ConFormer, Swin-Transformer, etc., with CNN models being the preferred choice.

[0046] Given the significant differences in scale, morphology, and spatial location of cerebral hemorrhage lesions, single-scale feature extraction often fails to fully capture their varied representational information, easily leading to the omission of small lesions or inaccurate characterization of large lesion boundaries, thus limiting the performance ceiling of segmentation models. To overcome this limitation, a multi-scale contextual feature fusion mechanism is designed and adopted in one optional implementation. Specifically, the feature extraction module extracts feature maps at multiple scales from brain image sample T, scales them to the same preset size, and then concatenates them along the channel dimension to obtain feature map F. CNN models, ConFormer, Swin-Transformer, etc., can be used for multi-scale feature extraction of brain images; no specific limitation is made here. In this implementation, a CNN model is preferred.

[0047] In one optional implementation, the diagnostic module is used to perform max pooling (or average pooling) and sigmoid activation operations on the logistic prediction score map in sequence to obtain the result map; when there are pixel values ​​in the result map that are greater than a preset threshold (generally 0.5), the corresponding pixel points belong to the foreground region, that is, the brain image is a brain hemorrhage image.

[0048] In one alternative implementation, after obtaining the class activation map, the localization module scales the resolution to the size of the brain image sample T and binarizes it to obtain the detection result of the brain hemorrhage lesion region in the brain image sample T.

[0049] This embodiment utilizes a diagnostic module and a localization module to achieve integrated and accurate "classification-localization" detection. While obtaining the classification results of the cerebral hemorrhage image, it adaptively generates a high-resolution class activation map and directly obtains the pixel-level localization results of the lesion through post-processing. This process deeply integrates the global image classification task with the local lesion localization task, enabling the localization process to be achieved solely based on image-level labels without additional segmentation and annotation supervision. This significantly reduces the system's dependence on labeled data while ensuring the accuracy and efficiency of lesion localization.

[0050] 2) Use brain image samples that are classified as brain hemorrhage images in the current batch as simulated brain hemorrhage images; for each simulated brain hemorrhage image... : its feature map Based on its class activation graph Weighted average pooling is used to obtain the corresponding foreground features. Its feature map Based on its background class activation graph Weighted average pooling is used to obtain the corresponding background features. ;

[0051] In this embodiment, foreground features By analyzing feature maps Based on class activation graph Obtained by weighted average pooling; background features By analyzing feature maps Based on background class activation graph The foreground features are obtained by weighted average pooling. Specifically, in one optional implementation, the aforementioned foreground features... and background features They are respectively:

[0052]

[0053]

[0054] Where H and W are feature maps respectively. The number of rows and columns; Representation of feature map The pixel in the i-th row and j-th column; Representation of class activation graph The pixel in the i-th row and j-th column; Represents the background class activation graph The pixel in the i-th row and j-th column.

[0055] It should be noted that, in one optional implementation, the background class activation graph... Considering the complex and diverse background information in brain images, including various anatomical structures such as the skull and brain parenchyma, compressing all this information into a single background prototype would introduce significant noise and obfuscation. In another alternative implementation, the feature map... Based on background class activation graph Before performing weighted average pooling, the background class activation map is also processed. A random dropout operation (preferably 95% random dropout) was performed to obtain multiple background class prototypes with differentiated information within a training batch. At this point, the feature maps... Based on background class activation graph Background class activation graph during weighted average pooling To The result after performing a random dropout operation.

[0056] 3) For each simulated brain hemorrhage image Calculate its feature map The features at each pixel location are respectively related to and The similarity is used to obtain the corresponding foreground similarity map. Similarity map with background ;Will and After subtraction, the result is processed by the ReLU function and then normalized to obtain... The corresponding prototype class activation graph; specifically, The corresponding prototype class activation graph is calculated as follows: It should be noted that the above similarity can be cosine similarity, Euclidean distance, dot product, etc., with cosine similarity being the preferred method.

[0057] against The position of each pixel in the comparison and The region type corresponding to the larger similarity value at the corresponding position is taken as the region type to which the pixel belongs (i.e., when the similarity value is larger than the region type). The similarity value at the corresponding position is greater than If the similarity value at the corresponding position is equal to the value of the pixel, then the region type of that pixel is the foreground region; otherwise, the region type of that pixel is the background region. This allows us to obtain... The corresponding pseudo-label image ;based on Sure Corresponding Logistic Prediction Score Chart Background area in to obtain Maximum Logistic Prediction Score ;Will After subtracting the maximum logical prediction score from each pixel, the result is processed sequentially using the ReLU function and then normalized to obtain... The corresponding relative class activation graph; specifically, The corresponding relative class activation graph is calculated as follows: .

[0058] Considering that the logistic prediction score map generated by the classification network often suffers from significant smoothness issues during training relying solely on image-level labels, and that this smoothness manifests as insufficiently distinct response differences between lesion and background regions, this application constructs a relative class activation map as a key supervision signal. This map is generated by subtracting the maximum logistic prediction score of the background region from the logistic prediction score map of the brain image pixel by pixel, thereby sharpening the response differences between the foreground and background, significantly improving the response contrast between the lesion and the background, thus enhancing the saliency of localization, suppressing background noise, and greatly improving the robustness of the localization results to threshold selection.

[0059] 3) The brain hemorrhage lesion detection model is trained by minimizing the classification loss and consistency loss of the current batch;

[0060] For classification loss:

[0061] The classification loss represents the sum of the difference losses between the classification results and the corresponding classification labels of each brain image sample in the current batch.

[0062] It should be noted that the difference loss in classification loss can be measured using cross-entropy loss function, L2 loss function, cosine similarity loss, hinge loss, etc., and no specific method is used here.

[0063] Regarding consistency loss:

[0064] The consistency loss represents the sum of the difference losses between the prototype class activation maps and the relative class activation maps corresponding to each simulated brain hemorrhage image in the current batch. This consistency constraint mechanism, by forcibly aligning the relative class activation maps generated by image-level classification signal generation with the prototype class activation maps generated by prototype feature matching, provides a self-correcting capability for the weakly supervised training process. This mechanism effectively bridges the gap between global semantic judgment and local feature similarity, thereby suppressing localization noise or bias caused by the limitations of a single signal source. Through this cross-level consistency verification, the model is guided to learn more accurate and robust lesion representations, significantly improving the accuracy and generalization performance of localization results in complex anatomical contexts.

[0065] Preferably, in one optional implementation, while minimizing the classification loss and consistency loss of the current batch, the prototype difference loss of the current batch is also minimized; wherein, the prototype difference loss represents the foreground feature set. The set of foreground and background features in The sum of the similarities of each background feature in the image; This is the set of foreground features of each simulated brain hemorrhage image in the current batch; This is a set of background features of each simulated brain hemorrhage image in the current batch.

[0066] Regarding prototype difference loss: it should be noted that the foreground feature set... Foreground and background feature sets in The similarity of background features can be measured using methods such as cosine similarity, dot product, Euclidean distance, and Manhattan distance, without limitation here. Cosine similarity is preferred. In one optional implementation, the expression for the prototype difference loss is:

[0067]

[0068] Where N is the number of simulated brain hemorrhage images in the current batch.

[0069] This implementation employs an explicit difference maximization strategy for foreground and background features, actively widening the distance between different semantic prototypes in the feature space through a prototype difference loss function. This design enables the model to better distinguish lesions from diverse normal anatomical structures (such as cerebrospinal fluid, white matter, and gray matter), improving the discriminativeness of feature representations under weak supervision and enhancing the model's ability to differentiate lesions from complex background structures.

[0070] In one optional implementation, the total loss during training of the brain hemorrhage lesion detection model in each round of training is a weighted sum of the classification loss, prototype difference loss, and consistency loss for the current batch. In another optional implementation, the expression for the total loss is:

[0071]

[0072] in, Indicates the total loss. Represents classification loss. Indicates prototype difference loss. This indicates a loss of consistency.

[0073] In this implementation, prototype difference loss and consistency loss are combined to guide the model to learn feature representations with strong discriminative power. Prototype difference loss enhances the model's ability to identify lesion regions by actively widening the feature prototype distance between the lesion foreground and the complex background; consistency loss provides a self-correction mechanism for weakly supervised training by forcibly aligning localization maps generated by different signal sources, effectively suppressing localization noise. This application effectively improves the accuracy and reliability of cerebral hemorrhage lesion detection and localization under weak supervision settings using only image-level labels.

[0074] Figure 1 shows the training process of the cerebral hemorrhage lesion detection model provided in this embodiment; Figure 2 shows a schematic diagram of all relevant output results during the training process provided in this embodiment; Figure 3 shows the prediction probability distribution diagram without consistency loss provided in this embodiment; Figure 4 shows the prediction probability distribution diagram with consistency loss provided in this embodiment.

[0075] After training, the trained feature extraction module extracts feature maps from brain images. Subsequently, the diagnosis module generates a logistic prediction score map based on this feature map and provides image-level classification results for brain hemorrhage diagnosis. If an image is identified as a brain hemorrhage, the localization module is automatically activated, using the generated logistic prediction score map to further generate a high-resolution class activation map. After scaling and binarization, it finally outputs accurate pixel-level localization results for brain hemorrhage lesions. This embodiment integrates an end-to-end training strategy, incorporating mechanisms such as prototype difference maximization and consistency constraints. Therefore, even with training using only image-level labels, it can still achieve high-precision, automated detection and localization of brain hemorrhage lesions, effectively overcoming the prediction smoothness problem of traditional weakly supervised methods in the context of brain hemorrhage detection.

[0076] Example 2

[0077] This embodiment provides a method for detecting cerebral hemorrhage lesions, including:

[0078] The brain image to be detected is input into the brain hemorrhage lesion detection model to obtain the brain hemorrhage lesion detection results;

[0079] The cerebral hemorrhage lesion detection model was trained using the training method provided in Example 1 of this application. The cerebral hemorrhage lesion detection results include: the classification result of whether the brain image is a cerebral hemorrhage image, and the detection result of the cerebral hemorrhage lesion region when the brain image is a cerebral hemorrhage image.

[0080] The related technical solutions are the same as the training method provided in Embodiment 1 of this application, and are not limited here.

[0081] It should be noted that the brain images in the above embodiments are all based on brain CT images, but are not limited to brain CT images.

[0082] Example 3

[0083] This embodiment provides a brain hemorrhage lesion detection system, including: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the method provided in Embodiment 2 of this application.

[0084] The relevant technical solutions are the same as the brain hemorrhage lesion detection method provided in Embodiment 2 of this application, and are not limited here.

[0085] Example 4

[0086] This embodiment provides an electronic device, including: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the method provided in Embodiment 1 or Embodiment 2 of this application.

[0087] The related technical solutions are the same as the training method provided in Embodiment 1 of this application and the cerebral hemorrhage lesion detection method provided in Embodiment 2 of this application, and are not limited here.

[0088] Example 5

[0089] This embodiment also provides a computer-readable storage medium, which includes a stored computer program, wherein the computer program, when run by a processor, controls the device where the storage medium is located to execute the method provided in Embodiment 1 or Embodiment 2 of this application.

[0090] The related technical solutions are the same as the training method provided in Embodiment 1 of this application and the cerebral hemorrhage lesion detection method provided in Embodiment 2 of this application, and are not limited here.

[0091] Example 6

[0092] This embodiment also provides a computer program product, including a computer program / instruction that, when executed by a processor, implements the method provided in Embodiment 1 or Embodiment 2 of this application.

[0093] The related technical solutions are the same as the training method provided in Embodiment 1 of this application and the cerebral hemorrhage lesion detection method provided in Embodiment 2 of this application, and are not limited here.

[0094] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A training method for a cerebral hemorrhage lesion detection model, characterized in that, include: Brain image samples from the training set are input into a brain hemorrhage lesion detection model in batches for multiple rounds of training. The model includes a feature extraction module, a diagnosis module, and a localization module. Each round of training involves: inputting each brain image sample T from the current batch into the model; the feature extraction module extracts features from the brain image sample T to obtain a feature map F; the diagnosis module maps the feature map F to a logistic prediction score map L, thereby obtaining a classification result indicating whether T is a brain hemorrhage image; the localization module, when T is a brain hemorrhage image, processes L sequentially through a ReLU function and normalizes it to obtain a class activation map, thereby obtaining the detection result of the brain hemorrhage lesion region in T; brain image samples classified as brain hemorrhage images in the current batch are used as pseudo-brain hemorrhage images; for each pseudo-brain hemorrhage image... : its feature map Based on their class activation graphs respectively and background class activation image Weighted average pooling is used to obtain the corresponding foreground features. and background features ； ;calculate The features at each pixel location are respectively related to and The similarity is used to obtain the corresponding foreground similarity map. Similarity map with background ;Will and After subtraction, the result is processed by the ReLU function and then normalized to obtain... The corresponding prototype class activation graph; for The position of each pixel in the comparison and The similarity value at the corresponding position is used to determine the region type corresponding to the larger similarity value, thus obtaining the region type to which the pixel belongs. The corresponding pseudo-label image ;based on Sure Corresponding Logistic Prediction Score Chart Background area in ;Will pixel by pixel The difference between the maximum logistic prediction scores is then processed by the ReLU function and normalized to obtain the result. The corresponding relative class activation map; the model is trained by minimizing the classification loss and consistency loss of the current batch; the consistency loss represents the sum of the difference loss between the prototype class activation map and the relative class activation map corresponding to each simulated brain hemorrhage image in the current batch.

2. The training method according to claim 1, characterized in that, While minimizing the classification loss and consistency loss of the current batch, the prototype difference loss of the current batch is also minimized; wherein, the prototype difference loss represents the foreground feature set. The set of foreground and background features in The sum of the similarities of each background feature in the image; and These are the sets of foreground features and the sets of background features for each simulated brain hemorrhage image in the current batch.

3. The training method according to claim 2, characterized in that, The expression for the prototype difference loss is: Where N is the number of simulated brain hemorrhage images in the current batch.

4. The training method according to claim 1, characterized in that, The foreground features and the background features They are respectively: Where H and W are feature maps respectively. The number of rows and columns; Representation of feature map The pixel in the i-th row and j-th column; Representation of class activation graph The pixel in the i-th row and j-th column; Represents the background class activation graph The pixel in the i-th row and j-th column.

5. The training method according to any one of claims 1-4, characterized in that, In the feature map Based on background class activation graph Before performing weighted average pooling, it is also necessary to... Perform a random discard operation.

6. The training method according to any one of claims 1-4, characterized in that, The feature extraction module is used to extract feature maps at multiple scales from brain image sample T, scale them to the same preset size, and then stitch them together in the channel dimension to obtain feature map F.

7. A method for detecting cerebral hemorrhage lesions, characterized in that, include: The brain image to be detected is input into the brain hemorrhage lesion detection model to obtain the brain hemorrhage lesion detection result; wherein, the brain hemorrhage lesion detection model is trained using the training method described in any one of claims 1-6; the brain hemorrhage lesion detection result includes: the classification result of whether the brain image is a brain hemorrhage image, and the detection result of the brain hemorrhage lesion region when the brain image is a brain hemorrhage image.

8. A system for detecting cerebral hemorrhage lesions, characterized in that, include: A memory and a processor, the memory storing a computer program, the processor executing the computer program to perform the method of claim 7.

9. An electronic device, characterized in that, include: A memory and a processor, the memory storing a computer program, the processor executing the computer program to perform the method according to any one of claims 1-7.

10. A computer program product, characterized in that, Includes a computer program / instruction that, when executed by a processor, implements the method described in any one of claims 1-7.

Images