A breast cancer histological grading prediction method and device based on fusion multi-task learning

Abstract

The application relates to a breast cancer histological grading prediction method and system fusing multi-task learning, which comprises the following steps: collecting breast cancer multi-phase DCE-MRI image data, obtaining a multi-phase two-dimensional image block sequence; adopting a pre-determined fusion element strategy, obtaining breast cancer histological grading based on a multi-phase two-dimensional image block sequence and a pathological information joint prediction model; wherein the pathological information joint prediction model is obtained by multi-task learning training and comprises a parameter sharing module and a multi-task learning module; the parameter sharing module is used for bottom layer shared features; the multi-task learning module executes a breast cancer cell proliferation antigen Ki-67 expression level prediction auxiliary task and a breast cancer histological grading prediction main task in parallel. Compared with the prior art, the end-to-end multi-task learning framework designed in the application significantly improves the recognition accuracy and learning efficiency of breast cancer histological grading prediction by capturing the correlation between breast cancer pathological information.

Description

Technical Field

[0001] This invention relates to the field of computer processing and analysis technology based on medical images, and in particular to a method and device for predicting the histological grading of breast cancer by integrating multi-task learning. Background Technology

[0002] Accurate and reliable pathological information assessment plays a crucial role in achieving personalized diagnosis and treatment decisions for breast cancer. Histological grading, as an independent prognostic pathological indicator, comprehensively reflects the risk of early breast cancer cell differentiation, recurrence, and metastasis through multiple aspects such as the degree of duct formation, nuclear pleomorphism, and mitotic figures. Furthermore, the immunohistochemical marker Ki-67 cell proliferation index is a key factor distinguishing between the molecular subtypes of breast cancer, Luminal A and Luminal B HER2-negative subtypes. In clinical practice, the assessment of breast cancer tumor invasiveness mainly relies on local tissue biopsies. However, for professional pathologists, manually performing histological grading and Ki-67 immunohistochemical testing without any auxiliary tools to process pathological images is extremely time-consuming and laborious.

[0003] Compared to conventional mammography and ultrasound, dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) effectively reflects the anatomical morphology and functional dynamics of lesions and has become a routine method for breast cancer screening. With the development of computer-aided diagnostic / detection technologies, automatically identifying imaging biomarkers linking DCE-MRI to breast cancer pathology using artificial intelligence is currently a research hotspot. However, existing methods still have limitations: models are designed only for predicting single pathological features, failing to fully utilize the complementary information between these features. In practice, clinicians often need to consider multiple pathological and histological features simultaneously before making precise clinical decisions for patients with invasive breast cancer. For example, compared to low- to intermediate-grade invasive breast cancer patients, patients with poorly differentiated (Grade III) breast cancer are more likely to exhibit a non-luminal molecular subtype, indicating a higher risk of cancer cell spread; high expression levels of the cell proliferation index Ki-67 are also associated with high histological grade of breast cancer. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method and device for predicting the histological grading of breast cancer by integrating multi-task learning.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for predicting the histological grade of breast cancer by incorporating multi-task learning includes the following steps:

[0007] Multi-temporal DCE-MRI image data of breast cancer were acquired, and multi-temporal two-dimensional image block sequences were obtained after preprocessing.

[0008] A predetermined fusion meta-strategy is adopted to obtain the histological grade of breast cancer based on the multi-temporal two-dimensional image block sequence and the trained pathological information joint prediction model.

[0009] The pathological information joint prediction model is obtained by multi-task learning training. The pathological information joint prediction model includes a parameter sharing module and a multi-task learning module. The parameter sharing module is used to share features at the bottom level. The multi-task learning module performs the auxiliary task of predicting the expression level of breast cancer cell proliferation antigen Ki-67 and the main task of predicting the histological grade of breast cancer in parallel.

[0010] The preprocessing includes histogram equalization, multi-scale cropping, and intensity normalization.

[0011] The multi-temporal DCE-MRI imaging data for breast cancer includes pre-contrast dynamic contrast-enhanced scan sequences and multiple sets of post-contrast dynamic contrast-enhanced scan sequences.

[0012] The parameter sharing module uses a lightweight DenseNet as its backbone network, which is embedded with an efficient channel attention mechanism and / or a gated attention mechanism.

[0013] The multi-task learning module is built based on an improved multi-gated hybrid expert model, and different weights are assigned to the underlying shared features for different tasks.

[0014] When training the joint prediction model for the pathological information, the loss is based on batch Focal loss L. g Calculate the learning error of the auxiliary task, using a cost-sensitive loss L based on the class F1-score. k Calculate the learning error of the main task and express the model optimization loss function as:

[0015] L Total =λ1L g +λ2L k

[0016] Where λ1 and λ2 represent weighting factors that measure the impact of subtask loss.

[0017] The fusion meta-strategy includes an input-level fusion strategy, an intermediate feature-level fusion strategy, and a decision-level fusion strategy. The fusion meta-strategy with the best prediction accuracy is selected as the pre-determined fusion meta-strategy.

[0018] In the input-level fusion strategy, multi-temporal two-dimensional image block sequences are combined into multiple sets of scanning sequences in different ways. The image blocks of each set of scanning sequences are added pixel by pixel and cascaded, which serve as the input to the joint pathological information prediction model to obtain the histological grade of breast cancer.

[0019] In the intermediate feature-level fusion strategy, multi-temporal two-dimensional image block sequences are combined into multiple scan sequences in different ways. The underlying shared features of each scan sequence are obtained through the parameter sharing module. The underlying shared features are then combined and used as the input of the multi-task learning module to obtain the histological grade of breast cancer.

[0020] In the decision-level fusion strategy, multi-temporal two-dimensional image block sequences are combined into multiple sets of scanning sequences in different ways. Each set of scanning sequences is used as the input of the joint pathological information prediction model. The obtained prediction results are weighted and averaged to obtain the histological grade of breast cancer.

[0021] A breast cancer histological grading prediction system integrating multi-task learning includes:

[0022] The data acquisition device is used to acquire multi-temporal DCE-MRI image data of breast cancer, and obtain multi-temporal two-dimensional image block sequences after preprocessing;

[0023] A grading prediction device is used to obtain the histological grade of breast cancer based on the multi-temporal two-dimensional image block sequence and a trained pathological information joint prediction model by employing a predetermined fusion meta-strategy.

[0024] The pathological information joint prediction model is obtained by multi-task learning training. The pathological information joint prediction model includes a parameter sharing module and a multi-task learning module. The parameter sharing module is used to share features at the bottom level. The multi-task learning module performs the auxiliary task of predicting the expression level of breast cancer cell proliferation antigen Ki-67 and the main task of predicting the histological grade of breast cancer in parallel.

[0025] An electronic device, comprising:

[0026] One or more processors;

[0027] Memory; and

[0028] One or more programs stored in memory, the one or more programs including instructions for performing the breast cancer histological grading prediction method fused with multi-task learning as described above.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. The pathological information joint prediction model of this invention is obtained based on multi-task learning training. It complements the information between the prognostic factors of breast cancer pathology, histological grade and Ki-67 cell proliferation status, further improving the prediction effect of histological grade.

[0031] 2. This invention fully utilizes multi-temporal sequences of DCE-MRI as network input and designs an improved lightweight DenseNet as a parameter-sharing module to effectively extract high-throughput features, ensuring the accuracy of subsequent multi-task learning. The multi-task learning module is built based on an improved multi-gated hybrid expert model, designing a shared expert network and a task-specific network with their corresponding gating mechanisms, while also considering the commonalities and differences between task features, resulting in stronger model generalization ability.

[0032] 3. Based on the model's histological grading prediction performance under a single phase combination of DCE-MRI, this invention adaptively selects different fusion meta-strategies and makes full use of the information expressed by the features of each sequence, thereby obtaining the best grading prediction results. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the process of the present invention;

[0034] Figure 2 This is a schematic diagram illustrating the principle of multi-task learning fusion in the embodiment;

[0035] Figure 3 This is a schematic diagram of the parameter sharing module in the embodiment;

[0036] Figure 4 This is a schematic diagram of the multi-task learning module structure in the embodiment. Detailed Implementation

[0037] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0038] Example 1

[0039] This embodiment provides a method for predicting breast cancer histological grading by integrating multi-task learning, including the following steps: acquiring multi-temporal DCE-MRI image data of breast cancer, obtaining multi-temporal two-dimensional image block sequences after preprocessing, the preprocessing including histogram equalization, multi-scale cropping, and intensity normalization; and obtaining breast cancer histological grading based on a pre-determined fusion meta-strategy, using the multi-temporal two-dimensional image block sequences and a trained pathological information joint prediction model. The pathological information joint prediction model is obtained through multi-task learning training, and includes a parameter sharing module and a multi-task learning module. The parameter sharing module is used for sharing underlying features, while the multi-task learning module executes in parallel the auxiliary task of predicting the expression level of the breast cancer cell proliferation antigen Ki-67 and the main task of predicting breast cancer histological grading. Through the complementary information between the prognostic factors of breast cancer pathological information (historical grading) and Ki-67 cell proliferation status, the predictive effect of histological grading is improved.

[0040] In this embodiment, the acquired multi-temporal DCE-MRI image data of breast cancer includes the pre-contrast dynamic contrast-enhanced scanning sequence (TPs 1) and the post-contrast dynamic contrast-enhanced scanning sequences (TPs 2, TPs 3, and TPs 4). In this embodiment, a 3.0T dynamic contrast-enhanced e_THRIVE sequence was used for scanning. After intravenous injection of contrast agent, four phases were continuously repeated, denoted as TPs 1-TPs 4. Image preprocessing was performed on the MRI image data of each phase. To improve prediction accuracy, after obtaining the multi-temporal two-dimensional image block sequence, each was converted into a single-channel grayscale image and sequentially mapped to the RGB three channels of the network input. According to a pre-determined fusion meta-strategy, a combination of certain multi-temporal scanning sequences was input into the pathological information joint prediction model to obtain the histological grade of breast cancer.

[0041] I. Joint Prediction Model of Pathological Information

[0042] 1. Dataset Construction

[0043] Multi-temporal breast DCE-MRI image data were acquired, and two-dimensional image patches were obtained after preprocessing to construct a dataset. In this embodiment, the specific process of dataset construction includes:

[0044] ① Histogram equalization enhances image contrast.

[0045] ② Multi-scale cropping: After determining the image sequence segment containing the lesion area, the image block of the region of interest is cropped at multiple scales with the lesion location as the center, and the size of the square image block is readjusted to 224×224 pixels using a bilinear interpolation algorithm.

[0046] ③ An intensity normalization operation is used to unify the pixel values ​​of image blocks to [0,1] to facilitate network processing and analysis.

[0047] ④ After splitting the dataset using ten-fold cross-validation, the Python deep learning library Keras is used to perform real-time data augmentation operations (rotation, mirroring, scaling, etc.) only on the training set to improve the robustness of the model, while the validation set is left unprocessed.

[0048] The dataset contains DCE-MRI images and corresponding pathological label information. The pathological label information includes the histological grade of breast cancer and the proliferation status of Ki-67 cells. Pathologists scored breast cancer lesions according to the Nottingham Histological Grading System, including low-to-intermediate grades (Grade I, Grade II) and high grades (Grade III). Furthermore, referring to the Chinese Anti-Cancer Association Breast Cancer Diagnosis and Treatment Guidelines (2021 Edition), a 20% threshold was used to determine the Ki-67 proliferation index: Ki-67 ≥ 20% indicates high Ki-67 expression; Ki-67 ≤ 20% indicates low Ki-67 expression.

[0049] 2. Model Building

[0050] The pathological information joint prediction model constructed in this embodiment includes a parameter sharing module and a multi-task learning module, such as... Figure 1 and Figure 2 As shown.

[0051] like Figure 3 As shown, the parameter sharing module in this embodiment uses the lightweight DenseNet as its backbone network, with a convolutional layer growth rate k=12, and the number of Bottleneck units in the four Dense Block modules is set to [3, 6, 12, 8] respectively. Simultaneously, efficient channel attention (ECA) and gated attention (GAU) are inserted to ensure cross-channel interaction and effectively extract key feature information while avoiding the negative impact of accuracy loss caused by traditional SENet dimensionality reduction operations. Next, this module replaces the global average pooling layer with flattened layers and fully connected layers to aggregate and map the underlying shared features.

[0052] like Figure 4 As shown, the multi-task learning module in this embodiment utilizes an improved multi-gated hybrid expert model (CGC) to assign different weights to the shared feature vectors based on the differences between different sub-tasks. The principle is as follows:

[0053]

[0054]

[0055] g k (x)=w k (x)S k (x) (3)

[0056] y k (x)=L k (g k (x)) (4)

[0057] In this embodiment, only two sub-tasks are involved: breast cancer histological grading and Ki-67 expression level prediction. Therefore, k = {1, 2}. As shown in equation (1), for the k-th task, different Gate expert gating mechanisms w are first calculated. k (x); Secondly, the task-specific expert network in this embodiment... and shared expert network United S k (x), as shown in equation (2), calculate the top-level feature information g for different tasks. k (x), as in equation (3); finally, through the top layer L k (·) The output is the classification probability y. k (x), as in equation (4).

[0058] In this embodiment, using, as Figure 4 The two parallel fully connected layers and the classification layer shown serve as the top layer L. k (·) respectively achieve the auxiliary task of predicting the expression level of breast cancer cell proliferation antigen Ki-67 and the main task of predicting the histological grade of breast cancer.

[0059] In the multi-gated hybrid expert model (CGC) of this embodiment, considering the differences and interactions between different tasks, the shared underlying representation layer is divided into a specific task expert sub-network and a shared task expert sub-network. A multi-task gate expert gating is set to adaptively weight each expert sub-network to suppress the interference of redundant shared features on specific tasks.

[0060] 3. Model Training

[0061] Before training the joint prediction model for pathological information, the batch sample size was set to 64, and the initial learning rate was l. r =0.002, adaptively adjusting the learning rate and early stopping based on the validation set loss. L2 weight regularization is applied to each convolutional kernel with a coefficient of 0.005. During model training, the RMSprop optimizer is used to accelerate model parameter optimization. Batch Focal loss L... g Calculate the learning error for predicting the proliferation status of Ki-67 cells in the auxiliary task; use a cost-sensitive loss L based on class F1-score. k Calculate the learning error for the primary task of histological gradation prediction; jointly calculate the multi-task learning loss and optimize the model parameters:

[0062] L Total =λ1L g +λ2L k (5)

[0063]

[0064]

[0065] Where F(·)(F1-score) is a model performance evaluation metric composed of the harmonic mean of precision and recall, λ1 and λ2 represent weighting factors that measure the impact of sub-task loss, N is the batch size, c is the number of classes, and p g p k These are the model's predicted values ​​for the histological grading and Ki-67 expression level subtasks, respectively.

[0066] The initial training epochs were 120. After 10 epochs of training, the CFSL did not decrease, so the learning rate was reduced to 0.1% of its original value. After 40 fine-tuning iterations, L... Total If convergence has not yet occurred, the model training stops. Save the model weights that achieved the highest accuracy and lowest loss on the validation set.

[0067] II. Selection of Fusion Meta Strategy

[0068] In this embodiment, the preprocessed DCE-MRI multi-temporal sequence TPs1-4 image blocks are converted into single-channel grayscale images and sequentially mapped to the RGB three channels of the network input. The selection of the fusion meta-strategy is determined based on the impact analysis of the DCE-MRI time dimension on the performance of the breast cancer pathology information prediction model. In this embodiment, the four single-temporal sequences are arranged into three groups, named sequentially as the initial contrast enhancement sequence (TPs 1, 2, 3), the intermediate contrast enhancement sequence (TPs 1, 2, 4), and the final contrast enhancement sequence (TPs 2, 3, 4). (Reference) Figure 1 and Figure 2 As shown, by leveraging ensemble learning, the optimal fusion scheme is adaptively selected by comparing the histological grading performance of DCE-MRI phase combination sequences under different fusion meta-strategies.

[0069] In this embodiment, the fusion meta-strategy includes an input-level fusion method, an intermediate feature-level fusion method, and a decision-level fusion method, wherein:

[0070] In the input-level fusion method, the three sets of scanning sequences—the initial contrast-enhanced sequence (TPs 1,2,3), the intermediate contrast-enhanced sequence (TPs 1,2,4), and the final contrast-enhanced sequence (TPs 2,3,4)—are combined to form a DCE-MRI multi-temporal two-dimensional image block. The images are then added pixel by pixel and cascaded again as a single input to the joint pathological information prediction model to achieve preliminary prediction.

[0071] In the intermediate feature-level fusion method, the combination of different phases of DCE-MRI is used as the underlying shared features of the joint prediction model for input pathological information, and then connected to the same multi-task learning module to achieve preliminary prediction.

[0072] In the decision-level fusion method, the preliminary prediction results are obtained by using a weighted average fusion analysis of the pathological information obtained from the joint prediction model with different time-phase combinations of DCE-MRI as input.

[0073] By comparing and analyzing the preliminary prediction results obtained from the three fusion meta-strategies, the meta-strategy with the best prediction accuracy is selected as the optimal fusion scheme to achieve the final prediction.

[0074] III. Actual Forecast

[0075] After constructing and training a joint prediction model for obtaining pathological information and determining the optimal fusion scheme, the histological grading of real-time acquired multi-temporal DCE-MRI imaging data of breast cancer can be accurately predicted.

[0076] If the above methods are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0077] Example 2

[0078] This embodiment provides an electronic device including one or more processors, a memory, and one or more programs stored in the memory, the one or more programs including instructions for performing the breast cancer DCE-MRI histological grading prediction method fused with multi-task learning as described in Embodiment 1.

[0079] Example 3

[0080] This embodiment provides a breast cancer histological grading prediction system integrating multi-task learning, including a data acquisition device and a grading prediction device. The data acquisition device acquires multi-temporal DCE-MRI image data of breast cancer and obtains a multi-temporal two-dimensional image block sequence after preprocessing. The grading prediction device uses a pre-determined fusion meta-strategy to obtain the breast cancer histological grade based on the multi-temporal two-dimensional image block sequence and a trained pathological information joint prediction model. The rest is the same as in Embodiment 1.

[0081] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0082] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0083] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0084] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0085] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for predicting the histological grading of breast cancer by integrating multi-task learning, characterized in that, Includes the following steps: Multi-temporal DCE-MRI image data of breast cancer were acquired, and multi-temporal two-dimensional image block sequences were obtained after preprocessing; the multi-temporal DCE-MRI image data included dynamic contrast-enhanced pre-scanning sequence TPs 1 and multiple sets of dynamic contrast-enhanced post-scanning sequences TPs 2, TPs 3 and TPs 4; A predetermined fusion meta-strategy is adopted to obtain the histological grade of breast cancer based on the multi-temporal two-dimensional image block sequence and the trained pathological information joint prediction model. The pathological information joint prediction model is obtained by multi-task learning training. The pathological information joint prediction model includes a parameter sharing module and a multi-task learning module. The parameter sharing module is used for sharing features at the bottom level. The multi-task learning module executes the auxiliary task of predicting the expression level of breast cancer cell proliferation antigen Ki-67 and the main task of predicting the histological grade of breast cancer in parallel. The parameter sharing module uses a lightweight DenseNet as its backbone network, and the lightweight DenseNet is embedded with an efficient channel attention mechanism and / or a gated attention mechanism. The multi-task learning module is constructed based on an improved multi-gated hybrid expert model. The improved multi-gated hybrid expert model divides the low-level shared features output by the parameter sharing module into a specific task expert sub-network and a shared task expert sub-network, and sets corresponding multi-task gating for the main task and the auxiliary task respectively. For each task in the main task and the auxiliary task, the specific task expert sub-network and the shared task expert sub-network are adaptively weighted and then jointly combined through the multi-task gating corresponding to the task to obtain the top-level feature information of the task. Then, the top-level classification layer of the task outputs the prediction probability of the task, thereby assigning different weights to the low-level shared features for different tasks to suppress the interference of redundant shared features on specific tasks. When training the joint prediction model for pathological information, the learning error of the auxiliary task is calculated based on the batch Focal loss Lg, and the learning error of the main task is calculated based on the cost-sensitive loss Lk of the class F1-score. The model optimization loss function is expressed as: L_Total=λ1·Lg+λ2·Lk, where λ1 and λ2 represent the weighting factors that measure the impact of the sub-task loss. The fusion meta-strategy includes an input-level fusion strategy, an intermediate feature-level fusion strategy, and a decision-level fusion strategy. The multi-temporal two-dimensional image block sequences are arranged and combined according to the contrast enhancement time order into three sets of scanning sequences: initial contrast enhancement sequences TPs 1, 2, 3; intermediate contrast enhancement sequences TPs 1, 2, 4; and final contrast enhancement sequences TPs 2, 3, 4. In the input-level fusion strategy, image blocks of the same dimension in the three sets of scanning sequences are added pixel-by-pixel and concatenated along the channel dimension, serving as the input to the joint pathological information prediction model. In the intermediate feature-level fusion strategy, the underlying shared features of each set of scanning sequences are obtained through the parameter sharing module, and the underlying shared features of each set are concatenated along the feature dimension, serving as the input to the multi-task learning module. In the decision-level fusion strategy, each set of scanning sequences is used as the input to the joint pathological information prediction model, and the corresponding histological grade prediction probability distribution is obtained. Multiple prediction probability distributions are then weighted and averaged. The histological grade prediction accuracy of the above three fusion meta-strategies under different combinations of scanning sequences is compared, and the one with the best prediction accuracy is selected as the pre-determined fusion meta-strategy.

2. The breast cancer histological grading prediction method integrating multi-task learning according to claim 1, characterized in that, The preprocessing includes histogram equalization, multi-scale cropping, and intensity normalization.

3. A breast cancer histological grading prediction system integrating multi-task learning, characterized in that, include: The data acquisition device is used to acquire multi-temporal DCE-MRI image data of breast cancer, and obtain multi-temporal two-dimensional image block sequences after preprocessing; the multi-temporal DCE-MRI image data includes dynamic contrast enhancement pre-scanning sequence TPs 1 and multiple sets of dynamic contrast enhancement post-scanning sequences TPs 2, TPs 3 and TPs 4; A grading prediction device is used to obtain the histological grade of breast cancer based on the multi-temporal two-dimensional image block sequence and a trained pathological information joint prediction model by employing a predetermined fusion meta-strategy. The pathological information joint prediction model is obtained by multi-task learning training. The pathological information joint prediction model includes a parameter sharing module and a multi-task learning module. The parameter sharing module is used for sharing features at the bottom level. The multi-task learning module executes the auxiliary task of predicting the expression level of breast cancer cell proliferation antigen Ki-67 and the main task of predicting the histological grade of breast cancer in parallel. The parameter sharing module uses a lightweight DenseNet as its backbone network, and the lightweight DenseNet is embedded with an efficient channel attention mechanism and / or a gated attention mechanism. The multi-task learning module is constructed based on an improved multi-gated hybrid expert model. The improved multi-gated hybrid expert model divides the low-level shared features output by the parameter sharing module into a specific task expert sub-network and a shared task expert sub-network, and sets corresponding multi-task gating for the main task and the auxiliary task respectively. For each task in the main task and the auxiliary task, the specific task expert sub-network and the shared task expert sub-network are adaptively weighted and then jointly combined through the multi-task gating corresponding to the task to obtain the top-level feature information of the task. Then, the top-level classification layer of the task outputs the prediction probability of the task, thereby assigning different weights to the low-level shared features for different tasks to suppress the interference of redundant shared features on specific tasks. When training the joint prediction model for the pathological information, batch Focal loss is used. Calculate the learning error of the auxiliary task, using a cost-sensitive loss based on the class F1-score. Calculate the learning error of the main task and express the model optimization loss function as follows: in, This represents the weighting factor used to measure the impact of sub-task losses; The fusion meta-strategy includes an input-level fusion strategy, an intermediate feature-level fusion strategy, and a decision-level fusion strategy. The multi-temporal two-dimensional image block sequences are arranged and combined according to the contrast enhancement time order into three sets of scanning sequences: initial contrast enhancement sequences TPs 1, 2, 3; intermediate contrast enhancement sequences TPs 1, 2, 4; and final contrast enhancement sequences TPs 2, 3, 4. In the input-level fusion strategy, image blocks of the same dimension in the three sets of scanning sequences are added pixel-by-pixel and concatenated along the channel dimension, serving as the input to the joint pathological information prediction model. In the intermediate feature-level fusion strategy, the underlying shared features of each set of scanning sequences are obtained through the parameter sharing module, and the underlying shared features of each set are concatenated along the feature dimension, serving as the input to the multi-task learning module. In the decision-level fusion strategy, each set of scanning sequences is used as the input to the joint pathological information prediction model, and the corresponding histological grade prediction probability distribution is obtained. Multiple prediction probability distributions are then weighted and averaged. The histological grade prediction accuracy of the above three fusion meta-strategies under different combinations of scanning sequences is compared, and the one with the best prediction accuracy is selected as the pre-determined fusion meta-strategy.

4. An electronic device, characterized in that, include: One or more processors; Memory; and One or more programs stored in memory, the one or more programs including instructions for performing the breast cancer histological grading prediction method fused with multi-task learning as described in any one of claims 1-2.

Images