Medical image analysis system and analysis method based on multi-task learning

Abstract

The invention discloses a medical image analysis system and analysis method based on multi-task learning, and the system comprises an offline preprocessing module which is used for carrying out the image block segmentation and growth mode classification of a pathological full-slice image, and generating a pathological statistical vector based on the proportion of all types of image blocks; the multi-task learning module is used for receiving the three-dimensional medical image and the pathological statistical vector, and training the three-dimensional medical image by using the pathological statistical vector as a supervision signal so as to obtain a prediction model for predicting a clinical treatment decision; and the prediction deployment module is used for outputting a prediction result of the clinical treatment decision only based on the three-dimensional medical image in the reasoning stage. A multi-task learning system without cross-modal pixel-level registration is constructed by introducing a pathological statistical vector generated offline as a supervision signal, so that the model can accurately predict a clinical treatment decision only by depending on a three-dimensional medical image in a reasoning stage, and the decision difficulty caused by lack of pathological information in an operation is effectively overcome.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis technology, and in particular to a medical image analysis system and method based on multi-task learning. Background Technology

[0002] In recent years, artificial intelligence technology has made significant progress in the field of medical image analysis, particularly demonstrating great potential in the non-invasive prediction of pathological features of lung adenocarcinoma using preoperative chest CT images. The new pathological grading system for lung adenocarcinoma proposed by the International Association for the Study of Lung Cancer (IASLC) in 2020 emphasizes refined grading based on multiple growth patterns (such as lepidic, acinar, papillary, solid, and micropapillary) and their proportions, guiding individualized surgical strategies. For example, well-differentiated tumors may undergo sublobar resection to preserve lung function, while poorly differentiated tumors require lobectomy and systematic lymph node dissection. However, this grading relies on postoperative paraffin pathology, making accurate intraoperative evaluation via frozen section pathology difficult, leading to a dilemma for surgeons regarding "information gaps" and "time gaps" in surgical decision-making. To address this issue, researchers have attempted to construct predictive models that combine CT and pathology data. However, existing methods generally suffer from two major bottlenecks: First, establishing a direct correlation between CT and pathology images often relies on pixel-level cross-modal spatial registration. Since CT data is three-dimensional volumetric data and pathology data is two-dimensional slices, their spatial structure, scale, and deformation are fundamentally different, making accurate registration virtually impossible in clinical practice. Second, abandoning registration and relying solely on simple image-level labeling for supervision results in the loss of crucial information about the microscopic heterogeneity within the tumor, limiting the model's discriminative ability. Therefore, there is an urgent need for a cross-modal supervised learning paradigm that can effectively utilize the rich histological information contained in whole-slice pathology images, eliminates the need for complex registration, and is suitable for clinical deployment, in order to achieve the goal of accurately predicting clinical treatment decisions based solely on conventional three-dimensional CT images. Summary of the Invention

[0003] In view of this, the present invention proposes a medical image analysis system and method based on multi-task learning, which can accurately predict clinical treatment decisions based solely on conventional 3D CT images. The present invention provides the following technical solution: A medical image analysis system based on multi-task learning includes: The offline preprocessing module is used to segment and classify the growth patterns of whole pathological slide images, and generate pathological statistical vectors based on the proportion of each category of image blocks. The multi-task learning module is communicatively connected to the offline preprocessing module. It is used to receive three-dimensional medical images and pathological statistical vectors, and to use the pathological statistical vectors as supervision signals to train the three-dimensional medical images to obtain a predictive model for predicting clinical treatment decisions. The prediction deployment module, coupled with the multi-task learning module, is used to output the prediction results of the clinical treatment decision based solely on the three-dimensional medical images during the inference phase.

[0004] Optionally, the multi-task learning module includes a shared feature encoder, a main task head, and an auxiliary task head; The main task head is used to predict clinical treatment decisions; The auxiliary task head is used to output the prediction result of the pathological statistical vector, and to perform regularization constraints on the shared feature encoder based on the distribution difference between the prediction result and the pathological statistical vector during the training process.

[0005] Optionally, the offline preprocessing module is further configured to: Before classifying growth patterns, the image blocks are subjected to background filtering and tissue type filtering. The background filtering removes blank areas using a color space threshold method. The tissue type filtering uses a binary classification model to divide image blocks into tumor-type image blocks and non-tumor-type image blocks, and retains only tumor-type image blocks for subsequent growth pattern classification.

[0006] Optionally, the classification of growth patterns includes five growth patterns of tumors: lepidic, acinar, papillary, solid, or micropapillary. The pathological statistical vector is a five-dimensional probability vector, where each component represents the proportion of the number of image blocks of the corresponding growth mode to the total number of valid image blocks, and the sum of each component is 1.

[0007] Optionally, the shared feature encoder is a 3DCNN encoder, whose input is a 128×128×128 three-dimensional medical image after window width and window level normalization, and includes in sequence: The initial convolutional block is used to extract low-level texture features; The four-level residual block sequence has an increasing number of channels and progressively downsampled spatial resolution. And a global average pooling layer, used to output a shared feature vector of fixed dimensions.

[0008] Optionally, the distribution difference is measured using KL divergence; The total loss function of the multi-task learning module is a weighted combination: , in, The binary cross-entropy loss for the main task, To assist in the task, KL divergence loss, These are preset weighting coefficients.

[0009] This invention further discloses a medical image analysis method based on multi-task learning, comprising: The pathological whole-section images are segmented into image blocks and classified by growth pattern, and pathological statistical vectors are generated based on the proportion of image blocks of each category. The system receives three-dimensional medical images and uses the pathological statistical vectors as supervision signals to train the three-dimensional medical images to obtain a predictive model for predicting clinical treatment decisions. During the inference phase, the prediction results for the clinical treatment decision are output through the prediction model based solely on the three-dimensional medical images.

[0010] Optionally, training the three-dimensional medical image using the pathological statistical vector as a supervisory signal includes: The main task and auxiliary task can be performed simultaneously by sharing a feature encoder; The main task is used to predict clinical treatment decisions based on the three-dimensional medical images. The auxiliary task is used to output the prediction result of the pathological statistical vector, and to apply regularization constraints to the shared feature encoder based on the distribution difference between the prediction result and the pathological statistical vector.

[0011] Optionally, the distribution difference is measured using KL divergence; The total loss function during training is a weighted combination: , in, The binary cross-entropy loss for the main task, To assist in the task, KL divergence loss, These are preset weighting coefficients.

[0012] The present invention further discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described method.

[0013] This invention constructs a medical image analysis system based on multi-task learning, introducing pathological statistical vectors generated offline from whole-slice pathological images as supervisory signals. This achieves efficient cross-modal supervised learning of 3D medical images without requiring pixel-level registration between CT and pathological images. During the training phase, the system uses these pathological statistical vectors to guide the model in learning image features related to tumor microscopic growth patterns. During the inference phase, it can accurately predict clinical treatment decisions using only conventional 3D medical images, effectively solving the information and time lag problems caused by the difficulty in finely grading intraoperative frozen section pathology. This not only bypasses the technical bottleneck of multimodal alignment and significantly improves the accuracy and biological interpretability of treatment decision prediction, but also achieves a complex training yet lightweight deployment clinical architecture, providing reliable intelligent decision support for personalized treatment that can be directly embedded into existing diagnostic and treatment processes. Attached Figure Description

[0014] For illustrative purposes and not limiting, the present invention will now be described in conjunction with embodiments and accompanying drawings, wherein: Figure 1 This is a schematic diagram of the structure of the medical image analysis system according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the medical image analysis method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the electronic device in an embodiment of the present invention; Figure 4 This is a schematic diagram of the 3DCNN encoder structure in an embodiment of the present invention. Detailed Implementation

[0015] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application.

[0016] It should be noted that, where there is no conflict, the embodiments and features of the embodiments in this application can be combined with each other. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0017] refer to Figure 1This embodiment discloses a medical image analysis system based on multi-task learning, including an offline preprocessing module 11, a multi-task learning module 12, and a prediction deployment module 13. This embodiment takes lung tumor analysis as an example. In one or more embodiments, this embodiment can be applied to different disease types. The following is a detailed description: The offline preprocessing module 11 is used to segment and classify the growth patterns of the whole pathological slide images, and generate pathological statistical vectors based on the proportion of each type of image block.

[0018] In this embodiment, the offline preprocessing module 11 is used to automatically process the whole-slide image (WSI) to generate pathological statistical vectors for subsequent multi-task learning. Specifically, the module first divides the entire WSI into multiple image tiles of the same size according to preset rules, for example, each tile is 256×256 pixels; during the segmentation process, an overlap sampling strategy is used to ensure the spatial coverage integrity of key tumor regions, and each tile is quality-screened, retaining only high-quality image tiles with a tissue content of more than 50% and no obvious blurring or artifacts.

[0019] Subsequently, background filtering and tissue type filtering are applied sequentially to the selected tiles. Background filtering involves converting the image to HSV or LAB color space and setting brightness and saturation thresholds to remove blank areas. Tissue type filtering uses a pre-trained binary classification model ("tumor" vs. "non-tumor") to retain only tiles identified as tumor tissue, excluding non-target components such as normal lung tissue, interstitium, and necrosis. After these two filtering steps, the total number of valid tumor tiles is denoted as N.

[0020] Next, a 2D CNN classifier (such as ResNet-50 or DenseNet-121) fine-tuned on a large-scale pathological image dataset was used to classify the growth pattern of each tumor tile. The classification categories strictly correspond to the five tissue growth patterns of lung adenocarcinoma: lepidic, acinar, papillary, solid, and micropapillary. Regarding the categories, it should be noted that in one or more implementations, the number of categories C is not limited and can be set according to different disease types. For each category... The number of tiles predicted to be in this category is counted as [number]. Based on this, pathological statistical vectors The The formula for calculating each component is: ,in, Indicate category The proportion, ranging from [0,1], results in a five-dimensional probability vector for the final pathological statistics: Since the denominator is uniformly based on the total number of effective tumor tiles N, that is: This vector naturally satisfies the condition that the sum of its components is 1, fully representing the heterogeneous composition of the patient's tumor at the microscopic level, and serving as an auxiliary supervisory signal for the subsequent multi-task learning model. The entire preprocessing process is completed offline before model training and does not participate in the inference stage, thus ensuring the efficiency of system deployment and clinical applicability.

[0021] The multi-task learning module 12 is communicatively connected to the offline preprocessing module. It is used to receive three-dimensional medical images and pathological statistical vectors, and to use the pathological statistical vectors as supervision signals to train the three-dimensional medical images to obtain a predictive model for predicting clinical treatment decisions.

[0022] refer to Figure 4 In this embodiment, the multi-task learning module 12 is communicatively connected to the offline preprocessing module 11, and is used to jointly process three-dimensional medical images and pathological statistical vectors during the training phase to construct a prediction model with both high accuracy and biological interpretability. This module first receives preprocessed three-dimensional chest CT scan data, with an input size uniformly set to 128×128×128 voxels, and has undergone window width and level adjustment and normalization to adapt to the model input requirements. Simultaneously, it receives a five-dimensional pathological statistical vector generated by the offline preprocessing module 11, corresponding to the same patient. This vector represents the distribution of image blocks representing five growth patterns in the lung adenocarcinoma tumor: lepidic, acinar, papillary, solid, and micropapillary. The core of the multi-task learning module 12 is a shared 3DCNN encoder, whose structure includes an initial convolutional block, a four-level residual block sequence, and a global average pooling layer. It progressively extracts multi-level CT features from local texture to high-level semantics, and finally outputs a 512-dimensional shared feature vector.

[0023] Specifically, it includes the following levels: Initial convolutional block: The input CT volume first passes through an exemplary 7×7×7 3D convolutional layer with a stride of 2 to quickly reduce the spatial resolution and extract low-level texture features; then it is sequentially passed through a BatchNorm3D layer, a ReLU activation function, and a Dropout3D layer for regularization; finally, it passes through a 3×3×3 max pooling layer (with a stride of 2) to compress the feature map size to 32×32×32.

[0024] Level 4 residual block sequence: Residual Block 1: The main path contains two 3D convolutional layers (32 to 64 channels), with BatchNorm3D and ReLU in the middle, and Dropout3D at the end; the shortcut connection directly adds the input to the output, keeping the feature map size unchanged at 32×32×32.

[0025] Residual block 2: The number of main path channels changes from 64 to 128. The spatial resolution of the feature map is downsampled to 16×16×16 through convolution operations.

[0026] Residual block 3: The number of main path channels increases from 128 to 256, and the spatial resolution is further downsampled to 8×8×8.

[0027] Residual block 4: The number of main path channels is reduced from 256 to 512, and the spatial resolution is finally reduced to 4×4×4; Each residual block contains batch normalization, ReLU activation, and Dropout3D layers to enhance training stability and generalization ability.

[0028] Feature aggregation layer: After the fourth-level residual block, a global average pooling layer is connected to compress the 4×4×4 feature map into a 1×1×1 scalar feature; then, it is flattened into a one-dimensional vector through the Flatten layer, and finally outputs a 512-dimensional shared feature vector.

[0029] Based on this, the multi-task learning module 12 sets up two parallel task heads: The main task head uses a fully connected layer with Sigmoid activation to directly output the probability of whether a patient needs to undergo clinical treatment decisions such as lobectomy. The auxiliary task head uses a fully connected layer with Softmax activation to output a five-dimensional probability vector for predicting the aforementioned pathological statistical vector.

[0030] During training, the system simultaneously calculates two losses: the primary task uses binary cross-entropy loss to measure the deviation between the treatment prediction results and the actual clinical labels; the auxiliary task uses KL divergence loss to measure the difference in probability distribution between the predicted pathological statistical vector and the offline-generated actual pathological statistical vector. The total loss function is a weighted combination of the two: ,in For preset weighting coefficients, The binary cross-entropy loss for the main task, The KL divergence loss is used for the auxiliary task. While ensuring the performance of the main task, the auxiliary task applies regularization constraints to the shared feature encoder, guiding it to learn CT image representations highly correlated with the microscopic pathological structure of tumors. Through end-to-end backpropagation optimization, this module ultimately trains a predictive model that can accurately predict clinical treatment decisions and incorporates pathological knowledge, laying the foundation for subsequent non-invasive decision-making deployments relying solely on CT images.

[0031] The prediction deployment module 13, coupled to the multi-task learning module, is used to output the prediction results of the clinical treatment decision based solely on the three-dimensional medical images during the inference phase.

[0032] In this embodiment, the prediction deployment module 13 is coupled to the multi-task learning module 12 and is dedicated to the inference process during the clinical application phase. This module is independently invoked after the system training is completed. Its core feature is that it only requires input of conventional preoperative 3D chest CT images, without any pathological images or additional clinical data. Specifically, the model used during deployment is loaded from the weights obtained from the multi-task learning phase, but only the 3DCNN encoder and the main task head (i.e., the treatment prediction branch) are retained. The auxiliary task heads are removed during the deployment phase, thereby simplifying the model structure and reducing computational overhead. The input CT data first undergoes the same preprocessing workflow as the training phase, including window width and level adjustment (e.g., lung window: window width 1500 HU, window level −600 HU) and normalization. It is then uniformly resampled to a 128×128×128 voxel 3D tensor before being input into the model. The 3DCNN encoder, based on its pathobiologically significant feature extraction capabilities learned during training, automatically identifies image patterns related to lung adenocarcinoma growth patterns and high-risk factors, and outputs a treatment probability value between 0 and 1 through the main task head. This value indicates whether the patient should undergo a more aggressive surgical approach. Clinicians can interpret this probability based on preset decision thresholds or specific clinical contexts, thereby obtaining reliable, non-invasive, and individualized treatment recommendations preoperatively. The entire prediction process can be completed within seconds on a standard GPU server or an edge device integrated into a PACS system, without requiring full-slice pathology images, cross-modal pixel-level registration, or intraoperative waiting for frozen section pathology results. This truly achieves a clinically friendly intelligent decision support system that is complex to train but lightweight to deploy.

[0033] refer to Figure 2 This embodiment further discloses a medical image analysis method based on multi-task learning, applied to the above-mentioned medical image analysis system, including the following steps: S100: Performs image patch segmentation and growth pattern classification on whole-section pathological images, and generates pathological statistical vectors based on the proportion of image patches in each category. Specifically: S110: Automated processing of whole-slide images (WSI) to generate pathological statistical vectors for subsequent multi-task learning. Specifically, this module first segments the entire WSI into multiple uniformly sized image tiles according to preset rules, for example, each tile is 256×256 pixels. During segmentation, an overlap sampling strategy is used to ensure the spatial coverage integrity of key tumor regions, and each tile undergoes quality screening, retaining only high-quality image tiles with a tissue content exceeding 50% and no obvious blurring or artifacts.

[0034] S120: Perform background filtering and tissue type filtering sequentially on the selected tiles. Background filtering involves converting the image to HSV or LAB color space and setting brightness and saturation thresholds to remove blank areas. Tissue type filtering uses a pre-trained binary classification model ("tumor" vs. "non-tumor") to retain only tiles identified as tumor tissue, excluding non-target components such as normal lung tissue, interstitium, and necrosis. After these two filtering steps, the total number of valid tumor tiles is denoted as N.

[0035] S130: A 2D CNN classifier (such as ResNet-50 or DenseNet-121) fine-tuned on a large-scale pathological image dataset is used to classify the growth pattern of each tumor tile. The classification categories strictly correspond to the five tissue growth patterns of lung adenocarcinoma: extrinsic, acinar, papillary, solid, and micropapillary. Regarding the categories, it should be noted that in one or more implementations, the number of categories C is not limited and can be set according to different disease types. For each category... The number of tiles predicted to be in this category is counted as [number]. Based on this, pathological statistical vectors The The formula for calculating each component is: ,in, Indicate category The proportion, ranging from [0,1], results in a five-dimensional probability vector for the final pathological statistics: Since the denominator is uniformly based on the total number of effective tumor tiles N, that is: This vector naturally satisfies the condition that the sum of its components is 1, fully representing the heterogeneous composition of the patient's tumor at the microscopic level, and serving as an auxiliary supervisory signal for the subsequent multi-task learning model. The entire preprocessing process is completed offline before model training and does not participate in the inference stage.

[0036] S200: Receives three-dimensional medical images and uses the pathological statistical vectors as supervision signals to train the three-dimensional medical images to obtain a predictive model for predicting clinical treatment decisions. Specifically: Receives preoperative three-dimensional chest CT scan data obtained routinely in clinical practice. This CT data is uniformly preprocessed into a single-channel three-dimensional tensor of 128×128×128 voxels, and undergoes window width and window level adjustment (e.g., using a lung window: window width 1500HU, window level −600HU) and normalization to adapt to the model input requirements. Simultaneously, the system acquires the pathological statistical vectors generated by the aforementioned offline preprocessing module for the same patient corresponding to the CT data. This vector is obtained by statistically analyzing the proportion of image patches representing five types of lung adenocarcinoma growth patterns (lepidic, acinar, papillary, solid, and micropapillary) in whole-section pathological images, and satisfies the following conditions: .

[0037] The training process relies on a shared 3DCNN encoder, whose structure includes: initial 7×7×7 convolutional blocks combined with 3×3×3 max pooling to reduce the input spatial dimension to 32³; subsequently, a four-level residual block sequence is passed through, with the number of channels increasing from 32 to 512, and the spatial resolution gradually downsampled to 4³; finally, global average pooling and Flatten operations are performed to output a 512-dimensional shared feature vector. Based on this feature, the model branches out in parallel into two task heads: the main task head, consisting of fully connected layers and sigmoid activation, outputs the probability of whether a patient needs aggressive surgery such as lobectomy; the auxiliary task head, consisting of fully connected layers and softmax activation, outputs a five-dimensional prediction vector. , used to fit the true pathological statistical vector S.

[0038] During training, two losses are calculated simultaneously: the main task uses binary cross-entropy loss. The study measures the difference between predicted treatment outcomes and actual clinical decision labels; the auxiliary task uses KL divergence loss. Measure the predicted distribution The probability difference between the true distribution S and the actual distribution S. The total loss function is defined as a weighted combination: The weighting coefficient A coefficient of 0.7 was set to ensure the accuracy of clinical decision-making while imposing strong regularization constraints on the shared feature encoder through auxiliary tasks, guiding it to learn CT image representations highly correlated with the microscopic pathological structure of tumors. Through end-to-end backpropagation optimization, the entire model parameters were jointly updated, ultimately training a predictive model capable of accurately predicting clinical treatment decisions based solely on CT images. This model incorporates pathological knowledge during training but requires no pathological image input or cross-modal pixel-level registration during deployment.

[0039] S300: During the inference phase, the prediction results for the clinical treatment decision are output through the prediction model based solely on the three-dimensional medical images. Specifically, after completing multi-task training, the trained model is deployed and optimized to form a lightweight prediction model specifically for clinical inference. This prediction model retains only the 3DCNN encoder and the main task head, i.e., the treatment prediction branch. The auxiliary task heads are removed during the deployment phase, thereby reducing computational overhead and improving inference efficiency. In practical applications, the three-dimensional chest CT scan data routinely obtained before surgery is received from the patient, without any pathological whole-section images, frozen sections, or other additional modal information. The CT data first undergoes a preprocessing procedure completely consistent with the training phase: including adjusting the window width and level using lung windows and normalizing it to the [0,1] interval, followed by resampling into a single-channel three-dimensional tensor of 128×128×128 voxels, which is then input into the prediction model. Based on the pathobiological knowledge internalized during training, the 3DCNN encoder automatically extracts high-level semantic features highly correlated with the microscopic heterogeneity of lung adenocarcinoma from CT images and compresses them into a 512-dimensional shared feature vector. The main task head then processes this feature vector and outputs a continuous probability value between 0 and 1, indicating whether the patient should receive the appropriate treatment plan. Clinicians can interpret this probability based on preset thresholds or in conjunction with the specific condition, thereby obtaining personalized, non-invasive, and reliable treatment decision recommendations before surgery.

[0040] Figure 3 A schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention, such as... Figure 3 As shown, the electronic device 50 includes: a processor 501, a memory 502, and a bus 503; The processor 501 and the memory 502 communicate with each other via the bus 503; the processor 501 is used to call the program instructions in the memory 502 to execute the methods provided in the above-described embodiments.

[0041] This embodiment provides a non-transitory computer-readable storage medium that stores computer instructions that cause a computer to execute the methods provided in the above-described embodiments.

[0042] Those skilled in the art will understand that all or part of the steps of the above-described method implementation can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above-described method implementation. The aforementioned storage medium includes various storage media capable of storing program code, such as ROM, RAM, magnetic disk, or optical disk.

[0043] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0044] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of each embodiment or some parts of the embodiments.

[0045] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can occur depending on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A medical image analysis system based on multi-task learning, characterized in that, include: The offline preprocessing module is used to segment and classify the growth patterns of whole pathological slide images, and generate pathological statistical vectors based on the proportion of each category of image blocks. The multi-task learning module is communicatively connected to the offline preprocessing module. It is used to receive three-dimensional medical images and pathological statistical vectors, and to use the pathological statistical vectors as supervision signals to train the three-dimensional medical images to obtain a predictive model for predicting clinical treatment decisions. The prediction deployment module, coupled with the multi-task learning module, is used to output the prediction results of the clinical treatment decision based on the three-dimensional medical images during the inference phase.

2. The medical image analysis system according to claim 1, characterized in that, The multi-task learning module includes a shared feature encoder, a main task head, and an auxiliary task head; The main task head is used to predict clinical treatment decisions; The auxiliary task head is used to output the prediction result of the pathological statistical vector, and to perform regularization constraints on the shared feature encoder based on the distribution difference between the prediction result and the pathological statistical vector during the training process.

3. The medical image analysis system according to claim 1, characterized in that, The offline preprocessing module is also used for: Before classifying growth patterns, the image blocks are subjected to background filtering and tissue type filtering. The background filtering removes blank areas using a color space threshold method. The tissue type filtering uses a binary classification model to divide image blocks into tumor-type image blocks and non-tumor-type image blocks, and retains only tumor-type image blocks for subsequent growth pattern classification.

4. The medical image analysis system according to any one of claims 1-3, characterized in that, The classification of growth patterns includes five growth patterns of tumors: lepidic, acinar, papillary, solid, or micropapillary. The pathological statistical vector is a five-dimensional probability vector, where each component represents the proportion of the number of image blocks of the corresponding growth mode to the total number of valid image blocks, and the sum of each component is 1.

5. The medical image analysis system according to claim 2, characterized in that, The shared feature encoder is a 3DCNN encoder, whose input is a 128×128×128 three-dimensional medical image after window width and window level normalization, and includes in sequence: The initial convolutional block is used to extract low-level texture features; The four-level residual block sequence has an increasing number of channels and progressively downsampled spatial resolution. And a global average pooling layer, used to output a shared feature vector of fixed dimensions.

6. The medical image analysis system according to claim 2, characterized in that, The distribution differences are measured using KL divergence; The total loss function of the multi-task learning module is a weighted combination: ， in, The binary cross-entropy loss for the main task, To assist in the task, KL divergence loss, These are preset weighting coefficients.

7. A medical image analysis method based on multi-task learning, characterized in that, include: The pathological whole-section images are segmented into image blocks and classified by growth pattern, and pathological statistical vectors are generated based on the proportion of image blocks of each category. The system receives three-dimensional medical images and uses the pathological statistical vectors as supervision signals to train the three-dimensional medical images to obtain a predictive model for predicting clinical treatment decisions. During the inference phase, the prediction results for the clinical treatment decision are output through the prediction model based solely on the three-dimensional medical images.

8. The medical image analysis method according to claim 7, characterized in that, The step of using the pathological statistical vector as a supervisory signal to train the three-dimensional medical image includes: The main task and auxiliary task can be performed simultaneously by sharing a feature encoder; The main task is used to predict clinical treatment decisions based on the three-dimensional medical images. The auxiliary task is used to output the prediction result of the pathological statistical vector, and to apply regularization constraints to the shared feature encoder based on the distribution difference between the prediction result and the pathological statistical vector.

9. The medical image analysis method according to claim 7, characterized in that, The distribution differences are measured using KL divergence; The total loss function during training is a weighted combination: ， in, The binary cross-entropy loss for the main task, To assist in the task, KL divergence loss, These are preset weighting coefficients.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method of any one of claims 7-9.

Images