Classification and staging system and method for ovarian cancer

Abstract

The present application relates to a kind of ovarian cancer classification staging system and method, belong to medical image processing technical field.The problems of insufficient multi-image processing capability, poor multi-modal compatibility, fixed staging logic, insufficient explainability and clinical adaptability of existing ovarian cancer classification staging technology are solved.The ovarian cancer classification staging system of the present application includes data input module, pre-processing module, feature extraction module, dynamic aggregation module, double-branch reasoning module and result output module.The present application also provides ovarian cancer classification staging method.The ovarian cancer classification staging system and method simultaneously support CT and MRI input, and have high clinical adaptability;Using space-channel-sequence triple attention module, adaptive processing of indefinite number of full images is realized, and information utilization and classification accuracy are greatly improved;Using decision module, reasoning speed and logical rigor are improved;Standardized diagnosis unifies classification staging standard, and reduces subjective differences of different doctors.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to an ovarian cancer classification and staging system and method, which is particularly suitable for the automated analysis of gynecological tumor images. Background Technology

[0002] Ovarian cancer, the leading cause of death among malignant tumors of the female reproductive system, relies on the synergy of medical imaging techniques, including computed tomography (CT) / magnetic resonance imaging (MRI) and pathological examination, for early diagnosis. However, current medical imaging analysis technologies suffer from the following key technical deficiencies, making it difficult to meet clinical needs: 1. Insufficient multi-image processing capabilities: Current technologies for processing patient CT / MRI images generally employ a sampling strategy due to hardware computing power limitations, resulting in: (1) The image information loss rate is high, and small lesions are easily missed; (2) The sampling rules rely on human experience, are highly subjective, and have poor consistency in diagnosis among different doctors, with a Kappa coefficient of <0.7; (3) The inability to utilize the temporal / spatial correlation information of multiple images leads to low classification accuracy.

[0003] 2. Poor multimodal compatibility: CT and MRI have fundamentally different imaging characteristics. CT is based on bone and tissue density, while MRI has higher resolution for soft tissues. Most existing systems are designed as a single modality. (1) Systems that only support MRI cannot handle patients with circumcision rings; (2) Systems that only support CT have insufficient resolution for soft tissues; (3) Cross-modal features cannot be aligned, multimodal data cannot be used together, and clinical applicability is limited.

[0004] 3. Fixed installment logic: Current technologies uniformly implement a process of first classifying and then staging, without taking into account the differences in the biological behavior of ovarian tumors. (1) Staging results are still output for benign and borderline tumors, resulting in invalid calculations; (2) The staging model uses ordinary cross-entropy classification loss and does not consider the orderliness of FIGO staging (the progression of the disease from stage I to stage IV). (3) The classification and staging results are output independently without logical linkage. Doctors need to manually determine whether to accept the staging results.

[0005] 4. Insufficient interpretability and clinical fit: Most existing AI diagnostic systems are black box models: (1) Without attention visualization function, doctors cannot verify the lesion area that the model is focusing on, resulting in low trust level; (2) The model update cost is high (the entire network needs to be retrained when a new tumor subtype is added), making it difficult to adapt to the updates of clinical guidelines.

[0006] In summary, there is an urgent need for an ovarian cancer classification and staging technology that can adaptively process large amounts of multimodal images, optimize the staging process, and has clinical interpretability, in order to overcome the shortcomings of existing technologies. Summary of the Invention

[0007] In view of this, the present invention provides an ovarian cancer classification and staging system and method to address the problems of insufficient multi-image processing capabilities, poor multimodal compatibility, fixed staging logic, and insufficient interpretability and clinical adaptability of existing ovarian cancer classification and staging technologies.

[0008] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows.

[0009] In a first aspect, the present invention provides an ovarian cancer classification and staging system, comprising a data input module, a preprocessing module, a feature extraction module, a dynamic aggregation module, a two-branch reasoning module, and a result output module; The data input module automatically identifies the image modality by parsing the metadata identifier field in the DICOM metadata, obtains the patient's image sequence, and constructs a mapping table. The image sequence is one or both of CT image sequences and MRI image sequences. The mapping table includes at least the mapping relationship between histological subtype and biological behavior. The preprocessing module performs modality-specific preprocessing on each image acquired by the data input module to obtain a preprocessed image. The modality-specific preprocessing includes denoising, registration, normalization, and size unification. The feature extraction module extracts each preprocessed image obtained by the preprocessing module based on the 3D ResNet50 backbone network to obtain a backbone network feature map. The backbone network feature map is then enhanced sequentially through the spatial attention submodule and the channel attention submodule to obtain a 512-dimensional feature vector of a single backbone network feature map. The 512-dimensional feature vectors of all backbone network feature maps form a multi-image feature vector set. The dynamic aggregation module calculates the weight scores of the 512-dimensional feature vectors of each backbone network feature map obtained by the feature extraction module through the sequence attention submodule, and applies the softmax activation function to normalize all the obtained weight scores so that each weight score is in the range of [0,1] and the sum of all weight scores is 1. Finally, the multi-image feature vector set is weighted and aggregated based on the normalized weight scores to generate a fixed 512-dimensional patient-level global feature. The dual-branch reasoning module includes a classification branch module, a decision module, and a phased branch module; The classification branch module uses a 3-layer fully connected layer, BatchNorm and Dropout structure to process the patient-level global features, selects the one with the highest probability as the predicted classification, and outputs the predicted classification to the decision module. The decision module maps the predicted histological subtype to biological behavior based on a mapping table. When the biological behavior is malignant, the control staging branch module is activated. When the biological behavior is benign or borderline, the classification result is directly output to the result output module. The staging branch module uses a Transformer encoder and a fully connected layer structure to process patient-level global features, selects the one with the highest probability, obtains the FIGO I-IV staging results, and outputs them to the result output module. The result output module outputs the predicted classification and staging results.

[0010] Preferably, the image sequence is a CT image sequence. The preprocessing module performs modality-specific preprocessing on each CT image as follows: First, the CT image is subjected to 3D Gaussian filtering with a standard deviation of σ=1.5 and a convolution kernel of 3×3×3. Then, threshold segmentation is performed based on the ovarian HU value [-100,0]. Next, the HU value is truncated to [-1000, 400] and linearly normalized to [0,1]. Finally, the size is adjusted to 32×128×128 through trilinear interpolation to obtain the preprocessed image. The image sequence is an MRI image sequence. The preprocessing module performs modality-specific preprocessing on each MRI image as follows: First, the MRI image is denoised using nonlocal mean with a window size of 7×7×3; then, high signal regions are identified based on the T2-weighted sequence; then, the image is standardized by channel Z-score with mean = 0 and variance = 1; finally, the image size is adjusted to 32×128×128 using trilinear interpolation to obtain the preprocessed image.

[0011] Preferably, the feature extraction module extracts features from each preprocessed image obtained by the preprocessing module based on the 3D ResNet50 backbone network, and the process of obtaining the backbone network feature map is as follows: (1) Replace the first convolutional layer of the 3D ResNet50 backbone network with a modality-adaptive convolutional layer: The image sequence is a CT image sequence. First, the 3-channel CT image is mapped to a 64-channel feature map, and then a 3×3×3 convolution kernel is used with 1 pixel padding. The image sequence is an MRI image sequence, in which 3-channel MRI images are filled with a 3×3×3 convolution kernel and 1 pixel. (2) Add dilated convolutions to the 3rd and 4th layers of the 3D ResNet50 backbone network, with an expansion rate of 2, to increase the receptive field to 64×64×16 and obtain the backbone network feature map.

[0012] Preferably, the spatial attention submodule first performs max pooling and average pooling on the backbone network feature map along the channel dimension to obtain two first feature maps; then, the two obtained first feature maps are concatenated and 3D convolution is performed, wherein the 3D convolution kernel is 3×3×3, the pixel padding is 1, and it is compressed into 1 channel; then, the sigmoid activation function is used to obtain weight scores, each weight score is in the range [0,1] and the sum of all weight scores is 1; finally, it is multiplied element-wise with the backbone network feature map to obtain the spatial attention feature map; The channel attention submodule first compresses the spatial dimension of the spatial attention feature map to 1×1×1 using adaptive average pooling to obtain the second feature map; then, it uses two fully connected layers 512→128→512 to compress the dimensionality of the channel features of the second feature map; then, it uses the sigmoid activation function to obtain weight scores, each weight score being in the range [0,1] and the sum of all weight scores being 1; finally, it multiplies element-wise with the spatial attention feature map to obtain a 512-dimensional feature vector of a single backbone network feature map.

[0013] Preferably, the sequence attention submodule first maps the 512-dimensional feature vector of each backbone network feature map to a weight score through a linear layer, then applies the softmax activation function to the obtained weight scores, each weight score is in the range of [0,1], and the sum of all weight scores is 1. Finally, the multi-image feature vector set is aggregated according to the weight scores to generate a patient-level global feature vector. The linear layer has an input dimension of 512 and an output dimension of 1. The formula for calculating the weighted score is: ; In the formula, Indicates the first Weight scores of the 512-dimensional eigenvectors of the backbone network feature map. This represents the natural exponential function. This represents a linear layer mapping operation. Indicates the first The 512-dimensional feature vector of the backbone network feature map The 512-dimensional feature vector index representing the backbone network feature map, from 1 to... The 512-dimensional feature vectors of all backbone network feature maps traversed. This indicates the total number of images for a single patient; The weighted aggregation formula is as follows: ; In the formula, This represents the patient-level global feature vector with a dimension of 512.

[0014] More preferably, the result output module further includes a quality assessment unit, which assesses the quality of patient-level global features.

[0015] More preferably, the result output module generates a heatmap and a weight score distribution map using the weight scores obtained by the spatial attention submodule.

[0016] Preferably, the classification branch module uses a 3-layer fully connected layer of 512→256→128→32, and outputs the probabilities of the top three subtypes to the decision module.

[0017] Preferably, the staging branching module uses a Transformer encoder and a fully connected layer 512→4 to process the patient-level global features, and selects the one with the highest probability to obtain the FIGO stage I~IV staging results.

[0018] Preferably, the classification branching module also outputs the expected classifications of the second and third positions to the decision module; The result output module outputs the top three predicted classifications, their corresponding probabilities and staging results, as well as the classification confidence and staging confidence.

[0019] Preferably, the ovarian cancer classification and staging system further includes an optimization training module, which optimizes the parameters of the feature extraction module, dynamic aggregation module, classification branch module, and staging branch module. The loss function for the feature extraction module, dynamic aggregation module, and classification branch module is the cross-entropy loss function; The loss function for the phased branching module adopts the ordered logistic regression loss function, formula: ; In the formula, For the cumulative indicator variable of the installment label, For cumulative prediction probability, This is the total loss value. This indicates the serial number of FIGO's installments. =1, 2, 3 or 4, corresponding to Phase I, Phase II, Phase III and Phase IV respectively.

[0020] Secondly, the present invention also provides a method for classifying and staging ovarian cancer using the above-mentioned ovarian cancer classification and staging system, comprising the following steps: S1, the data input module automatically identifies the image modality by parsing the metadata identifier field in the DICOM metadata, obtains the patient's image sequence, and constructs a mapping table. The image sequence is one or both of CT image sequences and MRI image sequences. The mapping table includes at least histological subtypes and biological behaviors, and the mapping table includes at least the mapping relationship between histological subtypes and biological behaviors. S2, the preprocessing module performs modality-specific preprocessing on each image acquired by the data input module to obtain a preprocessed image. The modality-specific preprocessing includes denoising, registration, normalization and size unification. S3, the feature extraction module extracts each preprocessed image obtained by the preprocessing module based on the 3D ResNet50 backbone network to obtain backbone network feature maps, and enhances the backbone network feature maps in sequence through the spatial attention submodule and the channel attention submodule to obtain the 512-dimensional feature vector of a single backbone network feature map. The 512-dimensional feature vectors of all backbone network feature maps form a multi-image feature vector set. S4, the dynamic aggregation module calculates the weight scores of the 512-dimensional feature vectors of each backbone network feature map obtained by the feature extraction module through the sequence attention submodule, and applies the softmax activation function to normalize all the obtained weight scores so that each weight score is in the range of [0,1] and the sum of all weight scores is 1. Finally, the multi-image feature vector set is weighted and aggregated based on the normalized weight scores to generate a fixed 512-dimensional patient-level global feature. S5, the classification branch module uses a 3-layer fully connected layer, BatchNorm and Dropout structure to process the patient-level global features, takes the one with the highest probability as the predicted classification, and outputs the predicted classification to the decision module; The decision module maps the predicted histological subtype to biological behavior based on the mapping table. When the biological behavior is malignant, the control staging branch module is activated. When the biological behavior is benign or borderline, the classification result is directly output to the result output module. The staging branch module uses a Transformer encoder and a fully connected layer structure to process patient-level global features, selects the feature with the highest probability, obtains the FIGO I-IV staging results, and outputs them to the results output module. S6, the results output module outputs the predicted classification and staging results.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: The ovarian cancer classification and staging system and method of the present invention constructs a multimodal adaptation mechanism and a specific preprocessing strategy, while supporting CT and MRI input, achieving seamless compatibility between CT and MRI dual modalities, automatically identifying image modalities and matching the optimal processing flow, solving the clinical pain point that patients with IUDs cannot undergo MRI examinations, and its clinical adaptability is significantly better than that of single-modal systems.

[0022] The ovarian cancer classification and staging system and method of the present invention adopts a spatial-channel-sequence triple attention module, combined with a dynamic weighted aggregation algorithm, to achieve adaptive processing of an indefinite number of full images, thus solving the problem of information loss caused by traditional sampling.

[0023] The ovarian cancer classification and staging system and method of the present invention prevents the weight scores from being concentrated in a few images by calculating the entropy value of the sequence attention weight scores.

[0024] The ovarian cancer classification and staging system and method of the present invention breaks through the limitation of a fixed number of image inputs. Through a dynamic aggregation module, it can achieve adaptive processing of an indefinite number of images (800~1000 images / patient), which greatly improves information utilization and classification accuracy, solves the defect of information loss in the sampling of existing technologies, and supports rapid model updates when adding subtypes.

[0025] The ovarian cancer classification and staging system and method of this invention employs a decision module to construct a mapping table of histological subtypes and biological behaviors, and implements a staging activation decision mechanism. Benign / borderline tumors automatically skip the staging process, while malignant tumors are targeted to activate staging branches. At the same time, ordered logistic regression loss is used to adapt to the progressive characteristics of FIGO staging, solving the defects of redundant staging logic and ignoring the orderliness of staging in traditional techniques, improving reasoning speed, and processing time for a single patient is less than 5 seconds, meeting the real-time needs of clinical practice, and exhibiting high logical rigor.

[0026] The ovarian cancer classification and staging system and method of the present invention have strong interpretability and clinical adaptability, and can provide spatial attention heatmaps and weight score distribution maps, allowing doctors to intuitively understand the diagnostic basis of the model and improve their trust.

[0027] The ovarian cancer classification and staging system and method of the present invention eliminates the need for doctors to manually sample and convert image formats, thus shortening diagnosis time and improving diagnostic efficiency.

[0028] The ovarian cancer classification and staging system and method of this invention standardizes the diagnostic criteria and unifies the classification and staging standards, reducing subjective differences among different doctors.

[0029] The ovarian cancer classification and staging system and method of the present invention do not require high-end imaging equipment; ordinary CT / MRI can be used, which helps to improve the level of ovarian cancer diagnosis in primary hospitals. Attached Figure Description

[0030] Figure 1This is a schematic diagram of the ovarian cancer classification and staging system of the present invention.

[0031] Figure 2 This is a flowchart of the spatial attention submodule of the ovarian cancer classification and staging system of the present invention.

[0032] Figure 3 This is a flowchart of the channel attention submodule of the ovarian cancer classification and staging system of the present invention.

[0033] Figure 4 This is a flowchart of the dynamic aggregation module of the ovarian cancer classification and staging system of the present invention.

[0034] Figure 5 This is a flowchart of the two-branch reasoning module of the ovarian cancer classification and staging system of the present invention. Detailed Implementation

[0035] To further understand the present invention, preferred embodiments of the present invention are described below. However, it should be understood that these descriptions are only for further illustrating the features and advantages of the present invention, and not for limiting the scope of the claims of the present invention.

[0036] like Figures 1-5 As shown, the ovarian cancer classification and staging system of the present invention includes a data input module, a preprocessing module, a feature extraction module, a dynamic aggregation module, a two-branch reasoning module, and a result output module; The data input module automatically identifies image modalities by parsing the metadata identifier field (Modality) in the DICOM metadata, obtains the patient's image sequence, and constructs a mapping table. The image sequence is one or both of CT image sequences and MRI image sequences. The mapping table includes at least the mapping relationship between histological subtype and biological behavior. The preprocessing module performs modality-specific preprocessing on each image acquired by the data input module to obtain a preprocessed image. The modality-specific preprocessing includes denoising, registration, normalization, and size unification. The feature extraction module extracts features from each preprocessed image obtained by the preprocessing module based on the 3D ResNet50 backbone network, obtaining a backbone network feature map (batch, C, D, H, W, C=512, where batch represents the number of images loaded each time; C represents the image channel dimension; D represents the 3D image layer thickness dimension; H represents the 3D image vertical dimension; W represents the 3D image horizontal dimension). The backbone network feature map is then enhanced sequentially through the spatial attention submodule and the channel attention submodule to obtain a 512-dimensional feature vector of a single backbone network feature map. The 512-dimensional feature vectors of all backbone network feature maps form a multi-image feature vector set (batch, num_imgs (number of images per patient), 512). The dynamic aggregation module calculates the weight scores of the 512-dimensional feature vectors of each backbone network feature map obtained by the feature extraction module through the sequence attention submodule, and applies the softmax activation function to normalize all the obtained weight scores so that each weight score is in the range of [0,1] and the sum of all weight scores is 1. Finally, the multi-image feature vector set is weighted and aggregated based on the normalized weight scores to generate a fixed 512-dimensional patient-level global feature (batch, 512). The dual-branch reasoning module includes a classification branch module, a decision module, and a phased branch module; The classification branch module uses a 3-layer fully connected layer, BatchNorm and Dropout structure to process the patient-level global features, takes the one with the highest probability (pred_class) as the predicted classification, and outputs the predicted classification to the decision module; The decision module maps the predicted histological subtype to biological behavior based on a mapping table. When the biological behavior is malignant, the control staging branch module is activated. When the biological behavior is benign or borderline, the classification result is directly output to the result output module. The staging branch module uses a Transformer encoder and a fully connected layer structure to process the patient-level global features, selects the one with the highest probability (pred_class), obtains the FIGO I-IV staging results, and outputs them to the result output module; The result output module outputs the predicted classification and staging results.

[0037] In this invention, the input type of the data input module can be CT images and MRI images; preferably CT images: DICOM format, 3 channels, pixel value range -3000~3000HU; preferably MRI images: DICOM format, 3 channels (T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI sequence)), pixel value range 0~65535.

[0038] In this invention, the number of CT image sequences and MRI image sequences is preferably 800 to 1000 images each.

[0039] In this invention, the data input module supports a variable number of image inputs, preferably employing batch loading and dynamic memory allocation techniques to avoid GPU memory overflow. For example, in some implementations, an NVIDIA graphics card (RTX3090) with 24GB of video memory can be used to process 100 images in a single batch.

[0040] The mapping table of this invention is constructed based on clinical pathology consensus and clinical data statistics, and is not limited to a specific data table structure. It can support the configuration of different mapping tables according to clinical application scenarios. The mapping table includes at least the core fields: histological subtype and corresponding biological behavioral characteristics, and can be extended to associate other clinical subtyping indicators according to actual needs. For example, in some embodiments, 32 subtypes are used, and the mapping table is shown in Table 1.

[0041] Table 1 Mapping Table

[0042] In this invention, the preprocessing module employs a modality-specific preprocessing strategy, ensuring the comparability of images from different modalities and devices. Ovarian center alignment addresses image shift issues caused by patient positioning differences. A preferred approach for CT image processing is to first perform 3D Gaussian filtering on the CT images (using the standard deviation of the Gaussian function). =1.5, convolution kernel 3×3×3), metal artifact removal, based on morphological reconstruction; then threshold segmentation based on ovarian HU value [-100,0] to locate the ovarian center; then HU value truncated to [-1000, 400], linearly normalized to [0,1]; finally, the size is adjusted to 32×128×128 by trilinear interpolation to obtain the preprocessed image. A preferred MRI image processing scheme is as follows: first, non-local mean denoising (window size 7×7×3) and motion artifact correction (based on phase alignment) are performed on the MRI image; then, high signal regions are identified based on T2-weighted sequences, and the ovarian center is located by contour matching; then, channel-specific Z-score standardization (mean=0, variance=1); finally, the size is adjusted to 32×128×128 by trilinear interpolation to obtain the preprocessed image. It should be noted that in MRI T2-weighted sequences, the ovarian parenchyma and ovarian tumor regions appear as high-signal areas in the images. The automatic localization method based on contour matching is used, which does not rely on fixed numerical thresholds, but completes the center localization by identifying the relatively high signal features of the ovarian region on T2WI.

[0043] In this invention, the preferred feature extraction module extracts features from each preprocessed image obtained by the preprocessing module based on a 3D ResNet50 backbone network. The process of obtaining the backbone network feature map is as follows: (1) Replace the first convolutional layer of the 3D ResNet50 backbone network with a modality-adaptive convolutional layer: The image sequence is a CT image sequence. First, the 3-channel CT image is mapped to a 64-channel feature map, and then a 3×3×3 convolution kernel and 1 pixel padding are used. The image sequence is an MRI image sequence, in which 3-channel MRI images are filled with a 3×3×3 convolution kernel and 1 pixel. (2) Add dilated convolutions to the 3rd and 4th layers of the 3D ResNet50 backbone network, and set the dilation rate (convolution pixel spacing stride, dilation) to 2 to increase the receptive field to 64×64×16 (this size can cover the feature area of ​​ovarian tumors and achieve complete lesion feature capture) to obtain the backbone network feature map.

[0044] like Figure 2 As shown, the spatial attention submodule first performs max pooling and average pooling on the backbone network feature map along the channel dimension to obtain two first feature maps (batch, 1, D, H, W); then, the two first feature maps are concatenated and 3D convolution is performed. The 3D convolution kernel is 3×3×3, the pixel padding is 1, and it is compressed to 1 channel; then, the sigmoid activation function is used to obtain weight scores. Each weight score is in the range [0,1] and the sum of all weight scores is 1; finally, it is multiplied element-wise with the backbone network feature map to focus on the tumor region and obtain the spatial attention feature map (batch, 512, D, H, W).

[0045] like Figure 3 As shown, the channel attention submodule first compresses the spatial dimension of the spatial attention feature map to 1×1×1 through adaptive average pooling to obtain the second feature map (batch, 512, 1, 1, 1); then, it uses two fully connected layers 512→128→512 to compress the dimensionality of the channel features of the second feature map; then, it uses the sigmoid activation function to obtain weight scores, each weight score is in the range [0,1] and the sum of all weight scores is 1; finally, it multiplies the spatial attention feature map element by element to obtain the 512-dimensional feature vector of a single backbone network feature map.

[0046] like Figure 4 As shown, the dynamic aggregation module can solve the problem of inputting any number of images and output patient-level global features of fixed dimensions (batch, 512). The sequence attention submodule first maps the 512-dimensional feature vector of each backbone network feature map to a weight score through a linear layer 512→1. Then, it applies the softmax activation function to the obtained weight scores. Each weight score is in the range [0,1], and the sum of all weight scores is 1 (to ensure that images containing lesion information receive higher weights). Finally, it aggregates the multi-image feature vector set according to the weight scores to generate a patient-level global feature vector. To address the problem of the variable number of input images, it outputs a 512-dimensional feature vector, thus solving the problem of fixed input dimension for the subsequent inference module. The formula for calculating the weighted score is as follows: ; In the formula, Indicates the first Weight scores of the 512-dimensional eigenvectors of the backbone network feature map. This represents the natural exponential function. This represents a linear layer mapping operation. Indicates the first The 512-dimensional feature vector of the backbone network feature map The 512-dimensional feature vector index representing the backbone network feature map, from 1 to... The 512-dimensional feature vectors of all backbone network feature maps traversed. This indicates the total number of images for a single patient; The weighted aggregation formula is as follows: ; In the formula, This represents the patient-level global feature vector with a dimension of 512.

[0047] In this invention, the preferred classification branch module uses a 3-layer fully connected layer 512→256→128→32, BatchNorm and Dropout (0.5) structure to process the patient-level global features, outputting a 32-dimensional probability vector, and taking the one with the highest probability as the predicted classification, and outputting the predicted classification to the decision module; more preferably, the expected classification of the top three subtypes is taken.

[0048] In this invention, the preferred staging branch module uses a Transformer encoder (with parameters set to 1 layer, d_model=512, nhead=8) and a fully connected layer (512→4) structure to process the patient-level global features, and then takes the one with the highest probability to obtain the FIGO stage I~IV staging results.

[0049] The result output module of this invention can output the following data: (1) Basic information: Patient ID, examination type (CT / MRI), number of images, processing time; (2) Classification information: the top three predicted categories, the corresponding probabilities, the corresponding biological behaviors, and the classification confidence (the highest probability value is used as the stage confidence). (3) Staging information: corresponding predicted stage, stage confidence (the highest probability value is used as the stage confidence), and corresponding stage basis (such as tumor size, whether it invades extraovarian tissues, and whether there is distant metastasis).

[0050] In this invention, the preferred result output module can generate heatmaps and weight score distribution maps based on the weight scores obtained from the spatial attention submodule. This visually displays the tumor regions that the model focuses on, helping doctors verify the model's diagnostic basis and increasing clinical trust.

[0051] In this invention, the preferred result output module preferably also includes a quality assessment unit, which assesses the quality of patient-level global features and sets conditions for prompting doctors to review, thereby improving diagnostic reliability.

[0052] In this invention, it is preferable to further include an optimization training module, which optimizes the parameters of the feature extraction module, dynamic aggregation module, classification branch module, and staging branch module. The loss function for the feature extraction module, dynamic aggregation module, and classification branch module is the cross-entropy loss function; The loss function for the phased branching module adopts the ordered logistic regression loss function, formula: ; In the formula, For the cumulative indicator variable of the installment label, For cumulative prediction probability, This is the total loss value. This indicates the serial number of FIGO's installments. =1, 2, 3 or 4, corresponding to Phase I, Phase II, Phase III and Phase IV respectively.

[0053] The ovarian cancer classification and staging method of the ovarian cancer classification and staging system of the present invention includes the following steps: S1, the data input module automatically identifies the image modality by parsing the metadata identifier field in the DICOM metadata, obtains the patient's image sequence, and constructs a mapping table. The image sequence is one or both of CT image sequences and MRI image sequences. The mapping table includes at least histological subtypes and biological behaviors. S2, the preprocessing module performs modality-specific preprocessing on each image acquired by the data input module to obtain a preprocessed image. The modality-specific preprocessing includes denoising, registration, normalization and size unification. S3, the feature extraction module extracts each preprocessed image obtained by the preprocessing module based on the 3D ResNet50 backbone network to obtain backbone network feature maps, and enhances the backbone network feature maps in sequence through the spatial attention submodule and the channel attention submodule to obtain the 512-dimensional feature vector of a single backbone network feature map. The 512-dimensional feature vectors of all backbone network feature maps form a multi-image feature vector set. S4, the dynamic aggregation module calculates the weight scores of the 512-dimensional feature vectors of each backbone network feature map obtained by the feature extraction module through the sequence attention submodule, and applies the softmax activation function to normalize all the obtained weight scores so that each weight score is in the range of [0,1] and the sum of all weight scores is 1. Finally, the multi-image feature vector set is weighted and aggregated based on the normalized weight scores to generate a fixed 512-dimensional patient-level global feature. S5, the classification branch module uses a 3-layer fully connected layer, BatchNorm and Dropout structure to process the patient-level global features, takes the one with the highest probability as the predicted classification, and outputs the predicted classification to the decision module; The decision module maps the predicted histological subtype to biological behavior based on the mapping table. When the biological behavior is malignant, the control staging branch module is activated. When the biological behavior is benign or borderline, the classification result is directly output to the result output module. The staging branch module uses a Transformer encoder and a fully connected layer structure to process patient-level global features, selects the feature with the highest probability, obtains the FIGO I-IV staging results, and outputs them to the results output module. S6, the results output module outputs the predicted classification and staging results.

[0054] The ovarian cancer classification and staging method of the present invention can be trained and optimized through the following steps: S1, acquire image sequences of the ovarian region of patients with ovarian tumors and annotate them. The image sequences are one or both of CT image sequences and MRI image sequences. The annotation information includes histological classification, biological behavior and FIGO stage I-IV of malignant tumor. S2, divide the dataset into a training set and a validation set. The training set is used for training, and the validation set is used for validation. S3. Using training and validation sets, the parameters of the feature extraction module, dynamic aggregation module, classification branch module, and staging branch module are optimized and trained. A phased training strategy is adopted to gradually optimize the performance of the ovarian cancer classification method of the present invention and avoid gradient explosion in the early stage of training. The stages are shown in Table 2. Table 2. Phased Training Strategy

[0055] The loss function of the feature extraction module, the dynamic aggregation module, and the classification branch module is the cross-entropy loss function, which is a prior art technique. The loss function for the phased branching module adopts the ordered logistic regression loss function, formula: ; In the formula, For installment label, For cumulative prediction probability, This is the total loss value. This indicates the serial number of FIGO's installments. =1, 2, 3 or 4, corresponding to Phase I, Phase II, Phase III and Phase IV respectively.

[0056] In this invention, there are no special restrictions on the data source; any imaging data (CT and / or MRI) of ovarian tumor patients obtained through legal and compliant means is acceptable. The number of images per patient can be 800-1000. The training set and the validation set are typically allocated in a 4:1 ratio. For example, the training set consists of 960 cases (80%), including 400 CT cases and 560 MRI cases, covering all 32 subtypes and ensuring that the sample size for each subtype is ≥20 cases; the validation set consists of 240 cases (20%), including 50 CT cases and 70 MRI cases, with a subtype distribution similar to that of the training set. However, this is not the only possible approach.

[0057] In this invention, there are no special restrictions on the hardware and software environment in which it runs, such as: Hardware environment: Training server: GPU RTX3090 24GB VRAM, CPU i7-13700KF 16 cores, storage (2TB SSD + 4TB HDD). Inference client: Standard desktop computer (CPU i7-12700K, 32GB RAM, GPU RTX 3090 24GB), meeting the needs of real-time clinical inference; The software environment is shown in Table 3.

[0058] Table 3 Software Environment

[0059] By employing the aforementioned phased training strategy, the model's feature extraction, dynamic aggregation, and bi-branch inference performance can be optimized layer by layer, avoiding problems such as gradient explosion and overfitting, and ensuring stable network convergence. Through joint optimization of classification loss and staging ordered regression loss, the model can effectively learn the multimodal imaging features of ovarian tumors, achieving accurate joint diagnosis of histological classification and FIGO staging, meeting the requirements of reliability, real-time performance, and interpretability for clinical auxiliary diagnosis.

[0060] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the embodiments. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all embodiments here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. An ovarian cancer classification and staging system, characterized in that, It includes a data input module, a preprocessing module, a feature extraction module, a dynamic aggregation module, a two-branch inference module, and a result output module; The data input module automatically identifies the image modality by parsing the metadata identifier field in the DICOM metadata, obtains the patient's image sequence, and constructs a mapping table. The image sequence is one or both of CT image sequences and MRI image sequences. The mapping table includes at least the mapping relationship between histological subtype and biological behavior. The preprocessing module performs modality-specific preprocessing on each image acquired by the data input module to obtain a preprocessed image. The modality-specific preprocessing includes denoising, registration, normalization, and size unification. The feature extraction module extracts each preprocessed image obtained by the preprocessing module based on the 3D ResNet50 backbone network to obtain a backbone network feature map. The backbone network feature map is then enhanced sequentially through the spatial attention submodule and the channel attention submodule to obtain a 512-dimensional feature vector of a single backbone network feature map. The 512-dimensional feature vectors of all backbone network feature maps form a multi-image feature vector set. The dynamic aggregation module calculates the weight scores of the 512-dimensional feature vectors of each backbone network feature map obtained by the feature extraction module through the sequence attention submodule, and applies the softmax activation function to normalize all the obtained weight scores so that each weight score is in the range of [0,1] and the sum of all weight scores is 1. Finally, the multi-image feature vector set is weighted and aggregated based on the normalized weight scores to generate a fixed 512-dimensional patient-level global feature. The dual-branch reasoning module includes a classification branch module, a decision module, and a phased branch module; The classification branch module uses a 3-layer fully connected layer, BatchNorm and Dropout structure to process the patient-level global features, selects the one with the highest probability as the predicted classification, and outputs the predicted classification to the decision module. The decision module maps the predicted histological subtype to biological behavior based on a mapping table. When the biological behavior is malignant, the control staging branch module is activated. When the biological behavior is benign or borderline, the classification result is directly output to the result output module. The staging branch module uses a Transformer encoder and a fully connected layer structure to process patient-level global features, selects the one with the highest probability, obtains the FIGO I-IV staging results, and outputs them to the result output module. The result output module outputs the predicted classification and staging results.

2. The ovarian cancer classification and staging system according to claim 1, characterized in that, The image sequence is a CT image sequence. The preprocessing module performs modality-specific preprocessing on each CT image as follows: First, the CT image is subjected to 3D Gaussian filtering with a standard deviation of σ=1.5 and a convolution kernel of 3×3×3. Then, threshold segmentation is performed based on the ovarian HU value [-100,0]. Next, the HU value is truncated to [-1000, 400] and linearly normalized to [0,1]. Finally, the size is adjusted to 32×128×128 through trilinear interpolation to obtain the preprocessed image. The image sequence is an MRI image sequence. The preprocessing module performs modality-specific preprocessing on each MRI image as follows: First, the MRI image is denoised using nonlocal mean with a window size of 7×7×3; then, high signal regions are identified based on the T2-weighted sequence; then, the image is standardized by channel Z-score with mean = 0 and variance = 1; finally, the image size is adjusted to 32×128×128 using trilinear interpolation to obtain the preprocessed image.

3. The ovarian cancer classification and staging system according to claim 1, characterized in that, The feature extraction module extracts features from each preprocessed image obtained by the preprocessing module based on the 3DResNet50 backbone network. The process of obtaining the backbone network feature map is as follows: (1) Replace the first convolutional layer of the 3D ResNet50 backbone network with a modality-adaptive convolutional layer: The image sequence is a CT image sequence. First, the 3-channel CT image is mapped to a 64-channel feature map, and then a 3×3×3 convolution kernel is used with 1 pixel padding. The image sequence is an MRI image sequence, in which 3-channel MRI images are filled with a 3×3×3 convolution kernel and 1 pixel. (2) Add dilated convolutions to the 3rd and 4th layers of the 3D ResNet50 backbone network, with an expansion rate of 2, to increase the receptive field to 64×64×16 and obtain the backbone network feature map.

4. The ovarian cancer classification and staging system according to claim 1, characterized in that, The spatial attention submodule first performs max pooling and average pooling on the backbone network feature map along the channel dimension to obtain two first feature maps; then, the two first feature maps are concatenated and 3D convolution is performed, with a 3×3×3 kernel, pixel padding of 1, and compression to 1 channel; then, the sigmoid activation function is used to obtain weight scores, each weight score is in the range [0,1] and the sum of all weight scores is 1; finally, it is multiplied element-wise with the backbone network feature map to obtain the spatial attention feature map. The channel attention submodule first compresses the spatial dimension of the spatial attention feature map to 1×1×1 using adaptive average pooling to obtain the second feature map; then, it uses two fully connected layers 512→128→512 to compress the dimensionality of the channel features of the second feature map; then, it uses the sigmoid activation function to obtain weight scores, each weight score being in the range [0,1] and the sum of all weight scores being 1; finally, it multiplies element-wise with the spatial attention feature map to obtain a 512-dimensional feature vector of a single backbone network feature map.

5. The ovarian cancer classification and staging system according to claim 1, characterized in that, The sequence attention submodule first maps the 512-dimensional feature vector of each backbone network feature map to a weight score through a linear layer, then applies the softmax activation function to the obtained weight scores. Each weight score is in the range of [0,1], and the sum of all weight scores is 1. Finally, the multi-image feature vector set is aggregated according to the weight scores to generate a patient-level global feature vector. The linear layer has an input dimension of 512 and an output dimension of 1. The formula for calculating the weighted score is: ； In the formula, Indicates the first Weight scores of the 512-dimensional eigenvectors of the backbone network feature map. This represents the natural exponential function. This represents a linear layer mapping operation. Indicates the first The 512-dimensional feature vector of the backbone network feature map The 512-dimensional feature vector index representing the backbone network feature map, from 1 to... The 512-dimensional feature vectors of all backbone network feature maps traversed. This indicates the total number of images for a single patient; The weighted aggregation formula is as follows: ； In the formula, This represents the patient-level global feature vector with a dimension of 512.

6. The ovarian cancer classification and staging system according to claim 5, characterized in that, It possesses one or more of the following characteristics: The result output module also includes a quality assessment unit, which assesses the quality of patient-level global features; The result output module generates a heatmap and a weight score distribution map using the weight scores obtained from the spatial attention submodule.

7. The ovarian cancer classification and staging system according to claim 1, characterized in that, The classification branch module uses a 3-layer fully connected layer of 512→256→128→32, and outputs the probabilities of the first three subtypes to the decision module. The staging branch module uses a Transformer encoder and a fully connected layer 512→4 to process the patient-level global features, and selects the feature with the highest probability to obtain the FIGO stage I~IV staging results.

8. The ovarian cancer classification and staging system according to claim 1, characterized in that, The classification branch module also outputs the expected classifications of the second and third positions to the decision module; The result output module outputs the top three predicted classifications, their corresponding probabilities and staging results, as well as the classification confidence and staging confidence.

9. The ovarian cancer classification and staging system according to claim 1, characterized in that, The ovarian cancer classification and staging system also includes an optimization training module, which optimizes the parameters of the feature extraction module, dynamic aggregation module, classification branch module, and staging branch module. The loss function for the feature extraction module, dynamic aggregation module, and classification branch module is the cross-entropy loss function; The loss function for the phased branching module adopts the ordered logistic regression loss function, formula: ； In the formula, For the cumulative indicator variable of the installment label, For cumulative prediction probability, This is the total loss value. This indicates the serial number of FIGO's installments. =1, 2, 3 or 4, corresponding to Phase I, Phase II, Phase III and Phase IV respectively.

10. The ovarian cancer classification and staging method of the ovarian cancer classification and staging system according to any one of claims 1 to 9, characterized in that, Includes the following steps: S1, the data input module automatically identifies the image modality by parsing the metadata identifier field in the DICOM metadata, obtains the patient's image sequence, and constructs a mapping table. The image sequence is one or both of CT image sequences and MRI image sequences. The mapping table includes at least histological subtypes and biological behaviors, and the mapping table includes at least the mapping relationship between histological subtypes and biological behaviors. S2, the preprocessing module performs modality-specific preprocessing on each image acquired by the data input module to obtain a preprocessed image. The modality-specific preprocessing includes denoising, registration, normalization and size unification. S3, the feature extraction module extracts each preprocessed image obtained by the preprocessing module based on the 3D ResNet50 backbone network to obtain backbone network feature maps, and enhances the backbone network feature maps in sequence through the spatial attention submodule and the channel attention submodule to obtain the 512-dimensional feature vector of a single backbone network feature map. The 512-dimensional feature vectors of all backbone network feature maps form a multi-image feature vector set. S4, the dynamic aggregation module calculates the weight scores of the 512-dimensional feature vectors of each backbone network feature map obtained by the feature extraction module through the sequence attention submodule, and applies the softmax activation function to normalize all the obtained weight scores so that each weight score is in the range of [0,1] and the sum of all weight scores is 1. Finally, the multi-image feature vector set is weighted and aggregated based on the normalized weight scores to generate a fixed 512-dimensional patient-level global feature. S5, the classification branch module uses a 3-layer fully connected layer, BatchNorm and Dropout structure to process the patient-level global features, takes the one with the highest probability as the predicted classification, and outputs the predicted classification to the decision module; The decision module maps the predicted histological subtype to biological behavior based on the mapping table. When the biological behavior is malignant, the control staging branch module is activated. When the biological behavior is benign or borderline, the classification result is directly output to the result output module. The staging branch module uses a Transformer encoder and a fully connected layer structure to process patient-level global features, selects the feature with the highest probability, obtains the FIGO I-IV staging results, and outputs them to the results output module. S6, the results output module outputs the predicted classification and staging results.

Images