Noninvasive tumor invasiveness evaluation system based on magnetic resonance image optimization

Abstract

The invention relates to the technical field of medical image intelligent analysis, and discloses a noninvasive tumor invasiveness evaluation system based on magnetic resonance image optimization. Comprising a data receiving module used for receiving conventional magnetic resonance image data of a user; the image processing module is used for preprocessing the conventional magnetic resonance image data to obtain preprocessed image data; the model obtaining module is used for obtaining the sample data with the pathological label and performing supervised learning on the sample data with the pathological label to obtain a tumor magnetic resonance image analysis model; inputting the preprocessed image data into the tumor magnetic resonance image analysis model, extracting imaging evidence, and predicting a corresponding pathological grading result; and the report generation module is used for integrating a visual interpretation technology based on the pathological grading result, clearly presenting a key image region concerned by the model during judgment, and generating a tumor invasiveness evaluation report.

Description

Technical Field

[0001] This application relates to the field of intelligent medical image analysis technology, and in particular to a non-invasive tumor invasiveness assessment system based on magnetic resonance imaging optimization. Background Technology

[0002] The clinical diagnosis of prostate cancer currently relies primarily on blood PSA testing, ultrasound, and palpation. Mainstream guidelines recommend further diagnostic procedures based on magnetic resonance imaging (MRI). However, this system suffers from two major flaws: first, diagnosis is highly dependent on physician subjective experience, leading to low consistency in judgment among different physicians; second, distinguishing high-scoring tumor subtypes is difficult, often resulting in excessive biopsies to avoid missed diagnoses. Existing MRI-based assisted diagnostic solutions largely remain at the level of traditional image feature analysis, lacking in-depth mining and intelligent processing of multimodal MRI data. For example, traditional methods fail to effectively address the impact of sample size differences on model performance—with large sample sizes, it's difficult to efficiently utilize the generalization ability of pre-trained models; with insufficient sample sizes, there's a lack of unsupervised learning mechanisms to improve the model's understanding of the intrinsic structure of medical images. Furthermore, existing technologies do not integrate clinical indicators (such as PSA values ​​and patient history) with image features, resulting in significant limitations of single-image analysis and a lack of interpretability techniques, making it difficult to intuitively present the model's decision-making basis and affecting clinical credibility. Summary of the Invention

[0003] This application provides a non-invasive tumor invasiveness assessment system based on magnetic resonance imaging optimization, which aims to solve the following problem.

[0004] In a first aspect, this application provides a non-invasive tumor invasiveness assessment system based on magnetic resonance imaging optimization, comprising:

[0005] The data receiving module is used to receive the user's conventional magnetic resonance imaging data, which includes T1-weighted image data, T2-weighted image data, diffusion-weighted imaging data, and dynamic contrast-enhanced image data.

[0006] The image processing module is used to preprocess the conventional magnetic resonance image data to obtain preprocessed image data; determine whether a disease-specific database with a large sample size exists; if it exists, obtain a base model pre-trained on natural images or a large number of other medical images, and input the preprocessed image data into the base model for fine-tuning to obtain a first model adapted to the tumor magnetic resonance image analysis task; if it does not exist, utilize the characteristics of the preprocessed image data itself for unsupervised pre-training, and by partially occluding the image and then having the base model predict, enable the base model to learn to understand the internal structure of the medical image to obtain a second model;

[0007] The model acquisition module is used to acquire sample data with pathological labels, input the sample data with pathological labels into the first model or the second model for supervised learning, and obtain a tumor magnetic resonance imaging analysis model; input the preprocessed image data into the tumor magnetic resonance imaging analysis model, extract image evidence, and predict the corresponding pathological grading results;

[0008] The report generation module is used to integrate visualization interpretation technology based on the pathological grading results to clearly present the key image regions that the model focuses on when making judgments. The key image regions include tumor edges, internal necrotic areas, or abnormal enhancement areas. Based on the spatial distribution of the key image regions and the corresponding key regions of tumor cell distribution and stromal infiltration in the postoperative pathological sections, a tumor invasiveness assessment report is generated.

[0009] In some embodiments, the preprocessing of the conventional magnetic resonance image data to obtain preprocessed image data includes: removing noise from the conventional magnetic resonance image data by performing an image denoising algorithm, normalizing the image grayscale values ​​using a standardization method, and aligning image data from different sequences to the same coordinate system based on a spatial registration algorithm to generate preprocessed image data.

[0010] In some embodiments, determining whether a disease-specific database with a large sample size exists includes: querying a preset database sample size threshold to determine whether the number of magnetic resonance imaging samples with pathological labels contained in the disease-specific database exceeds the threshold; if it exceeds the threshold, it is determined that a disease-specific database with a large sample size exists; otherwise, it is determined that it does not exist.

[0011] In some embodiments, the step of obtaining a pre-trained base model on natural images or a large number of other medical images, and inputting the preprocessed image data into the base model for fine-tuning to obtain a first model adapted to the tumor magnetic resonance image analysis task includes: obtaining a base model based on a convolutional neural network, retaining the convolutional layer parameters of the base model, replacing or adjusting the fully connected layer structure, inputting the preprocessed image data into the adjusted base model, and iteratively training the fully connected layer parameters through a transfer learning algorithm until the loss function value of the base model on the tumor magnetic resonance image feature extraction task converges to obtain the first model.

[0012] In some embodiments, if the aforementioned feature is not present, unsupervised pre-training is performed using the characteristics of the preprocessed image data itself. By partially occluding the image and then having the base model predict its features, the base model learns to understand the internal structure of the medical image, thus obtaining a second model. This includes: randomly selecting a local region in the preprocessed image data for occlusion, generating input data with an occlusion mask, inputting the input data into the base model, training the base model to predict the image features of the occluded region using a self-supervised learning algorithm, updating the model parameters based on the difference between the prediction results and the original image features, until the base model can accurately reconstruct the image structure of the unoccluded region, thus obtaining the second model.

[0013] In some embodiments, the step of inputting the pathologically labeled sample data into the first model or the second model for supervised learning to obtain a tumor magnetic resonance imaging analysis model includes: dividing the pathologically labeled sample data into a training set and a validation set; inputting the preprocessed image data and its corresponding pathological grading labels from the training set into the first model or the second model; optimizing the model parameters using a backpropagation algorithm; evaluating the accuracy of the model on the pathological grading task using a validation set; repeating the training and validation process until the accuracy of the first model or the second model on the validation set reaches a preset threshold, thereby obtaining the tumor magnetic resonance imaging analysis model.

[0014] In some embodiments, the step of inputting the preprocessed image data into the tumor magnetic resonance imaging analysis model to extract image evidence and predict the corresponding pathological grading results includes: extracting features from the preprocessed image data through the convolutional layer of the tumor magnetic resonance imaging analysis model to generate a feature map containing tumor edges, internal textures, and signal intensity; classifying the feature map through a fully connected layer; and outputting the corresponding pathological grading prediction results.

[0015] In some embodiments, the step of integrating visualization interpretation technology based on the pathological grading results to clearly present the key image regions that the model focuses on when making judgments includes: analyzing the decision-making process of the tumor magnetic resonance imaging analysis model using a gradient-weighted class activation mapping algorithm, generating a heat map reflecting the regions of interest of the model, overlaying the heat map onto the original magnetic resonance image, marking the key image regions of the tumor edge and internal necrotic areas or abnormal enhancement areas, and displaying them in a visual form.

[0016] In some embodiments, generating a tumor invasiveness assessment report based on the spatial distribution of the key image region and the corresponding key regions of tumor cell distribution and stromal infiltration in the postoperative pathological section includes: matching the spatial coordinates of the key image region with the spatial coordinates of the tumor cell distribution region and the stromal infiltration region marked by immunohistochemical staining in the postoperative pathological section, calculating the spatial overlap rate of the tumor cell distribution region and the stromal infiltration region, generating a quantitative assessment report on the extent and degree of tumor invasion based on the overlap rate, and attaching a key region comparison diagram.

[0017] In some embodiments, the system further includes: a clinical data fusion module for acquiring the patient's clinical indicator data, including blood PSA value, patient age, and family history of tumors; the model acquisition module is further used to perform feature fusion of the clinical indicator data and the preprocessed image data, and train the tumor magnetic resonance imaging analysis model through a multiple learning algorithm, so that the tumor magnetic resonance imaging analysis model can combine imaging features and clinical features to predict pathological grading.

[0018] This application achieves non-invasive prediction of pathological grading through deep feature extraction from multimodal MRI data, avoiding the invasive risks and over-examination associated with traditional biopsies. A standardized analysis process based on machine learning models reduces diagnostic bias caused by differences in physician subjective experience and improves diagnostic consistency across different centers. By dynamically switching between transfer learning and self-supervised pre-training strategies, it effectively addresses the issue of sample size differences. With large samples, pre-trained models improve efficiency, while with small samples, unsupervised learning enhances model robustness. Multi-learning integrating clinical indicators and imaging features, combined with individual patient differences (such as PSA values ​​and age), improves predictive accuracy. Visual interpretation technology clearly identifies key imaging regions of interest to the model (such as tumor margins and necrotic areas), and these regions spatially align with key areas in pathological sections, enhancing clinicians' trust in AI decision-making and assisting in preoperative planning and biopsy site determination.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic block diagram of a non-invasive tumor invasiveness assessment system based on magnetic resonance imaging optimization provided in one embodiment of this application;

[0022] Figure 2 This is a schematic diagram of the principle of a non-invasive tumor invasiveness assessment system based on magnetic resonance imaging optimization provided in one embodiment of this application.

[0023] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0026] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0027] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0028] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0029] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0030] The clinical diagnosis of prostate cancer currently relies primarily on blood PSA testing, ultrasound, and palpation. Mainstream guidelines recommend further diagnostic procedures based on magnetic resonance imaging (MRI). However, this system suffers from two major flaws: first, diagnosis is highly dependent on physician subjective experience, leading to low consistency in judgment among different physicians; second, distinguishing high-scoring tumor subtypes is difficult, often resulting in excessive biopsies to avoid missed diagnoses. Existing MRI-based assisted diagnostic solutions largely remain at the level of traditional image feature analysis, lacking in-depth mining and intelligent processing of multimodal MRI data. For example, traditional methods fail to effectively address the impact of sample size differences on model performance—with large sample sizes, it's difficult to efficiently utilize the generalization ability of pre-trained models; with insufficient sample sizes, there's a lack of unsupervised learning mechanisms to improve the model's understanding of the intrinsic structure of medical images. Furthermore, existing technologies do not integrate clinical indicators (such as PSA values ​​and patient history) with image features, resulting in significant limitations of single-image analysis and a lack of interpretability techniques, making it difficult to intuitively present the model's decision-making basis and affecting clinical credibility.

[0031] To solve the above problem, please refer to Figures 1 to 2 This application provides a non-invasive tumor invasiveness assessment system based on magnetic resonance imaging optimization, comprising: a data receiving module for receiving conventional magnetic resonance image data from a user, the conventional magnetic resonance image data including T1-weighted image data, T2-weighted image data, diffusion-weighted imaging data, and dynamic contrast-enhanced image data; an image processing module for preprocessing the conventional magnetic resonance image data to obtain preprocessed image data; determining whether a disease-specific database with a large sample size exists; if so, acquiring a pre-trained basic model on natural images or a large number of other medical images, and inputting the preprocessed image data into the basic model for fine-tuning to obtain a first model adapted to the tumor magnetic resonance image analysis task; if not, performing unsupervised pre-training using the characteristics of the preprocessed image data itself, by partially occluding the image and then training the basic model... The model prediction method enables the base model to learn to understand the internal structure of medical images, resulting in a second model. The model acquisition module acquires pathologically labeled sample data and inputs this data into either the first or second model for supervised learning, resulting in a tumor magnetic resonance imaging (MRI) image analysis model. The preprocessed image data is then input into the MRI image analysis model to extract image evidence and predict the corresponding pathological grading results. The report generation module integrates visualization and interpretation techniques based on the pathological grading results to clearly present the key image regions that the model focuses on when making its judgments. These key image regions include tumor margins, internal necrotic areas, or abnormally enhanced areas. Based on the spatial distribution of these key image regions and the corresponding key regions of tumor cell distribution and stromal infiltration in postoperative pathological sections, a tumor invasiveness assessment report is generated.

[0032] Specifically, this invention aims to address two major pain points in the current MRI diagnostic process for prostate cancer: high diagnostic subjectivity and the difficulty in distinguishing high-scoring tumor subtypes, leading to excessive biopsies. The system utilizes an innovative and intelligent data processing and analysis workflow to transform conventionally acquired multimodal MRI images into objective, accurate, and interpretable tumor invasiveness assessment reports.

[0033] The core innovation of the system lies in the construction of an adaptive, interpretable, and multi-dimensional integrated intelligent analysis engine. Instead of employing a single, fixed model, it dynamically selects the optimal training path based on available data resources (the size of the disease-specific database), ensuring performance and robustness across different application scenarios. Ultimately, the system not only outputs a prediction result but also reveals the decision-making basis of that result through visualization technology, thereby fostering effective collaboration with clinicians.

[0034] The data receiving module, serving as the system's data entry point, is responsible for receiving raw DICOM format data from Picture Archiving and Communication Systems (PACS) or other imaging equipment. This module receives and parses data via a standard DICOM network interface. The received data type is specifically multi-parameter magnetic resonance imaging of the prostate, including: T1-weighted image data: primarily used for observing anatomical structures and assessing hemorrhage; T2-weighted image data: the core sequence used to display the prostate's anatomical structure, tumor morphology, and location (typically appearing as a low-signal area); diffusion-weighted imaging data: by calculating the apparent diffusion coefficient (ADC) map, reflecting the degree of water molecule diffusion restriction within the tissue, a key indicator for assessing tumor cell density; and dynamic contrast-enhanced image data: by analyzing the time-signal intensity curves of contrast agent inflow and outflow, assessing tumor angiogenesis, an important marker of invasiveness.

[0035] The image processing module standardizes the original images and intelligently selects a model initialization strategy based on data conditions. The computer first performs the following operations on the received multi-sequence images: Image denoising: Advanced image processing algorithms (such as non-local mean filtering) are applied to reduce image noise and improve the signal-to-noise ratio, laying the foundation for subsequent feature extraction. Standardization / Normalization: Grayscale values ​​of images acquired from different devices and with different scanning parameters are standardized to eliminate non-biological differences and ensure the consistency of model input. Spatial registration: Images from different sequences, such as T2-weighted, DWI, and DCE, are precisely aligned to the same coordinate system to ensure that the subsequently extracted features are spatially matched.

[0036] The system queries a pre-defined database to determine if the number of MRI samples with pathological gold standard labels (such as Gleason scores) has reached a pre-defined threshold (e.g., several thousand cases). This determines the subsequent learning strategy to be adopted.

[0037] Path 1 (Large Sample Case - Transfer Learning and Fine-tuning): If the sample size is sufficient, the system will load a large convolutional neural network model (i.e., the "base model") pre-trained on a massive natural image dataset (such as ImageNet) or other large-scale medical images. The system retains the low-level convolutional layer parameters used for general feature extraction in the base model, replacing only the top-level classifier structure. Subsequently, pre-processed images from the disease-specific database are input into the model, and all or part of the network parameters are iteratively updated using a relatively small learning rate (i.e., "fine-tuning"). This process allows the model to quickly adapt from general image recognition capabilities and focus on tumor-related specific features in medical images.

[0038] Path Two (Small Sample Case - Self-Supervised Pre-training): If the sample size is insufficient, the system utilizes the MRI images themselves for "unsupervised" learning. Implementation: The computer randomly selects a rectangular or arbitrarily shaped region from the image and sets its pixel values ​​to zero or masks it (i.e., "occludes" it). The occluded image is then input into a model to be trained, which is then trained to predict the original content of the occluded portion.

[0039] Through a massive amount of such practice, the model is forced to learn to understand the normal structure and texture patterns of various tissues and organs in medical images, thereby gaining powerful feature representation capabilities and laying a solid foundation for subsequent supervised learning.

[0040] The model acquisition module is used to complete the final training of the model and to use the trained model to predict the pathological grading of images of new patients.

[0041] Regardless of which path the initial model is obtained through (model one or model two), it will undergo "guided" training with explicitly labeled pathological data in this stage. The system divides the labeled sample data (images and corresponding Gleason scores) into training and validation sets. The training set data is input into the model, which outputs a predicted score, and the error (loss function) between the predicted score and the true label is calculated. Through backpropagation, the system automatically adjusts millions of internal parameters of the model, making its predictions continuously approach the true pathological results. The validation set is used to monitor the training process and prevent overfitting. When the model's performance on the validation set (e.g., accuracy) is stable and reaches a preset standard, training is complete, resulting in the final "tumor magnetic resonance imaging analysis model". Preprocessed image data from new patients is input into this final model, and the model's convolutional layers automatically extract and combine image features from low to high levels. Finally, the classification layer outputs a specific pathological grading prediction result (e.g., low risk, intermediate risk, high risk).

[0042] The report generation module makes the model's "black box" decision-making process transparent and generates a comprehensive and intuitive clinical evaluation report. The system employs techniques such as Gradient Weighted Class Activation Mapping (Grad-CAM). This technique calculates the importance weight of each spatial location on the feature map when the model makes its final decision. Based on these weights, a "heatmap" is generated and overlaid on the original MR image. The highlighted areas (e.g., red) in the image represent the regions of evidence that the model considers most crucial.

[0043] Users can visually see whether the model focuses on the correct anatomical location, such as whether the tumor boundary is clear, whether there are non-enhancing areas indicating necrosis inside, or whether there are areas of abnormal rapid enhancement.

[0044] For patients who have already undergone surgery, the system can digitally compare the heatmap generated by the model with the postoperative pathological slides. Through spatial coordinate matching, the overlap rate between the model's area of ​​interest and the actual distribution of tumor cells and the area of ​​stromal infiltration on the pathological slides is calculated. This provides the most direct evidence for the model's effectiveness. The computer integrates all the above information and automatically generates a structured report. This report includes at least: a clear pathological grade prediction; a visualized heatmap highlighting key areas of interest; an invasiveness assessment based on anastomosis analysis (e.g., "high probability of tumor invasion beyond the prostate capsule"); and clinical decision support recommendations, such as "targeted biopsy recommended for the abnormally enhancing area on the left posterolateral side."

[0045] This system achieves accurate and non-invasive assessment of prostate cancer invasiveness through a tightly integrated and highly adaptive automated workflow. Its value lies not only in the accuracy of the results but also in the transparency and reliability of the process, making it a promising AI assistant for clinicians to optimize diagnostic processes and improve the patient experience.

[0046] In some embodiments, the preprocessing of the conventional magnetic resonance image data to obtain preprocessed image data includes: removing noise from the conventional magnetic resonance image data by performing an image denoising algorithm, normalizing the image grayscale values ​​using a standardization method, and aligning image data from different sequences to the same coordinate system based on a spatial registration algorithm to generate preprocessed image data.

[0047] This embodiment focuses on providing high-quality, standardized input data for subsequent artificial intelligence models. It details a complete preprocessing workflow for raw, multi-sequence magnetic resonance images, with three key operations: noise reduction, normalization, and spatial registration. The aim is to eliminate image differences caused by non-pathological factors, ensuring that the features learned by the model are entirely derived from real biological information, thereby improving the model's stability and generalization ability.

[0048] Image denoising inevitably introduces random noise during the acquisition of magnetic resonance images. This noise can interfere with the model's extraction of subtle features and even cause the model to learn incorrect noise patterns. The system invokes a pre-defined image processing algorithm library, such as nonlocal mean filtering or deep learning-based denoising models. These algorithms intelligently identify noise points in the image that are dissimilar to the statistical characteristics of surrounding tissues and smooth them, while preserving the true tissue boundaries and texture information to the greatest extent possible. For example, for ADC maps, denoising can more clearly reveal the true areas where water molecule diffusion is restricted.

[0049] Image grayscale normalization is crucial because the absolute signal intensity of images can vary depending on the MRI equipment used, or even the same equipment scanned at different times. This difference, while not caused by lesions, can significantly impact the model's judgment. The computer iterates through all pixels in the image, employing z-score normalization or minimum-maximum normalization. Taking z-score as an example, the algorithm calculates the mean (μ) and standard deviation (σ) of pixel values ​​across the entire image or a specific region of interest, then transforms each pixel value x as (x-μ) / σ. After this processing, the pixel values ​​of all input images will follow a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating device dependence and differences in brightness and contrast.

[0050] The user's T2WI, DWI, and DCE images were acquired at different time points, and even slight postural movements are unavoidable, even if the patient remains stationary. Without registration, features extracted from the same location in different sequences will be spatially misaligned, leading to information fusion errors. The system uses high-resolution T2-weighted images as the baseline reference space. Then, spatial registration algorithms such as rigid body or affine transformations are used to calculate spatial transformation parameters (including translation, rotation, and scaling) between DWI and DCE sequence images and T2WI. Subsequently, the DWI and DCE images are resampled and interpolated based on these parameters to ensure that each voxel is precisely aligned anatomically with the T2WI image. For example, this ensures that low-signal lesions seen on T2WI completely coincide with significantly low-signal regions at corresponding locations on the ADC map.

[0051] In some embodiments, determining whether a disease-specific database with a large sample size exists includes: querying a preset database sample size threshold to determine whether the number of magnetic resonance imaging samples with pathological labels contained in the disease-specific database exceeds the threshold; if it exceeds the threshold, it is determined that a disease-specific database with a large sample size exists; otherwise, it is determined that it does not exist.

[0052] This embodiment defines how the system automatically evaluates the scale of available data resources and makes key strategic decisions based on this. It transforms the originally subjective judgment of "data volume size" into objective and executable computer logic by presetting a quantitative sample size threshold. This is the first decision point for the entire system to achieve the "adaptive" ability.

[0053] During system deployment or initialization, a database sample size threshold (N) is jointly set by domain experts (such as radiologists and data scientists) based on experience. This threshold is usually based on machine learning theory to ensure that when the sample size is greater than N, fine-tuning through transfer learning is efficient and reliable. For example, N can be set to 3000 cases.

[0054] When a new model needs to be trained, the system automatically connects to a specified disease-specific database (for example, the local hospital's prostate cancer MRI research database). The system executes a database query command to count the total number of samples (M) that simultaneously have high-quality MRI images and verified postoperative pathological labels (such as Gleason scores).

[0055] The system compares the statistically obtained total number of samples M with the preset threshold N. If M ≥ N: The system determines that "there is a disease-specific database with a large sample size", and then automatically triggers the transfer learning and fine-tuning process described in Embodiment 3. If M < N: The system determines that "it does not exist", and then starts the self-supervised pre-training process. This judgment mechanism is the core for the entire system to flexibly adapt to different application scenarios (from large medical centers to small and medium-sized hospitals).

[0056] In some embodiments, obtaining a pre-trained base model on natural images or a large number of other medical images, inputting the preprocessed image data into the base model for fine-tuning to obtain a first model adapted to the tumor magnetic resonance image analysis task includes: obtaining a base model based on a convolutional neural network, retaining the convolutional layer parameters of the base model, replacing or adjusting the fully connected layer structure, inputting the preprocessed image data into the adjusted base model, and iteratively training the fully connected layer parameters through a transfer learning algorithm until the loss function value of the base model converges on the tumor magnetic resonance image feature extraction task to obtain the first model.

[0057] This embodiment details how to efficiently utilize an advanced model (base model) trained on large-scale general images through fine-tuning technology to quickly specialize it in the prostate MRI analysis task when data resources are sufficient. This method utilizes the powerful general feature extraction ability already possessed by the base model, avoiding the massive data and computing resources required for "training from scratch".

[0058] The system loads a mature, large-scale convolutional neural network as the base model from a model library, such as ResNet, DenseNet, or EfficientNet. These models are typically pre-trained on the ImageNet dataset, which contains millions of natural images, and have learned to recognize basic visual patterns such as edges, corners, and textures.

[0059] The computer retains all convolutional and pooling layers in the base model because these lower-level networks are typically responsible for extracting general, task-independent features. Simultaneously, the fully connected layers (classifiers) at the top of the model are removed or replaced because the original classifier was designed for the 1000 object categories of ImageNet. The system replaces it with a new, simpler classifier (possibly with only a few layers), whose output nodes correspond to the number of pathological grading categories in this task (e.g., low-risk, intermediate-risk, high-risk).

[0060] The preprocessed disease-specific database images generated in the above embodiments are input into this adjusted model. The system employs a transfer learning algorithm, using a larger learning rate in the new classifier part and a smaller learning rate in the inherited convolutional layer part for training. During training, the difference between the pathological grade predicted by the model and the true pathological label is calculated (i.e., the loss function value, such as cross-entropy loss). Through backpropagation and an optimizer (such as Adam), the system mainly updates the parameters of the new fully connected layer and makes small, fine-tuning adjustments to the parameters of the underlying convolutional layers to make their features more suitable for medical images. Training is continuously iterated, and the loss function value is monitored. When the loss function value no longer decreases significantly on the training set and tends to stabilize (converge), training stops. The model obtained at this point, specifically optimized for prostate MRI, is the first model.

[0061] In some embodiments, if the aforementioned feature is not present, unsupervised pre-training is performed using the characteristics of the preprocessed image data itself. By partially occluding the image and then having the base model predict its features, the base model learns to understand the internal structure of the medical image, thus obtaining a second model. This includes: randomly selecting a local region in the preprocessed image data for occlusion, generating input data with an occlusion mask, inputting the input data into the base model, training the base model to predict the image features of the occluded region using a self-supervised learning algorithm, updating the model parameters based on the difference between the prediction results and the original image features, until the base model can accurately reconstruct the image structure of the unoccluded region, thus obtaining the second model.

[0062] This embodiment offers an innovative solution for scenarios where the sample size of disease-specific databases is insufficient. It utilizes the medical image data itself, employing a "propose and answer" approach (i.e., self-supervised learning), allowing the model to autonomously learn the anatomical structures and pathophysiological patterns contained within the images without human annotation. The advantage of this method lies in its ability to fully utilize all collected data that may lack pathological labels.

[0063] The computer randomly selects a batch of data from the preprocessed image data. For each image, the system generates a random occlusion mask, which can be a rectangle, multiple small squares, or any shape. Then, based on the position of the mask, the pixel values ​​of the corresponding areas in the image are set to zero or random noise, thereby generating input data with the occlusion mask. This simulates situations in clinical settings where image quality is poor or some information is missing.

[0064] The task of training a self-supervised learning model is to "predict the original image content of the occluded portion". During training, the system compares the reconstructed image predicted by the model with the original unoccluded image and calculates the difference (such as mean squared error). Through a self-supervised learning algorithm (actually a specific loss function and optimization process), gradients are calculated and the parameters of the entire base model are updated. After extensive practice, in order to better complete the "puzzle" task, the model must internally learn to understand and memorize intrinsic structural knowledge such as "what tissue is typically around the prostate" and "what are the typical textures inside a tumor region".

[0065] When the model can accurately reconstruct the image structure of the unoccluded region, it means that it has acquired a powerful feature representation capability for data specific to this domain. At this point, training stops, and the resulting model is the second model.

[0066] In some embodiments, the step of inputting the pathologically labeled sample data into the first model or the second model for supervised learning to obtain a tumor magnetic resonance imaging analysis model includes: dividing the pathologically labeled sample data into a training set and a validation set; inputting the preprocessed image data and its corresponding pathological grading labels from the training set into the first model or the second model; optimizing the model parameters using a backpropagation algorithm; evaluating the accuracy of the model on the pathological grading task using a validation set; repeating the training and validation process until the accuracy of the first model or the second model on the validation set reaches a preset threshold, thereby obtaining the tumor magnetic resonance imaging analysis model.

[0067] This example describes how to perform final "refinement" training on the initial model obtained through the example using the ultimate gold standard—postoperative pathological grading labels. This is a crucial step in transforming the model from a "generalist" or "structural understander" into a "prostate cancer diagnostic expert."

[0068] The system randomly divides all its pathologically labeled sample data into training and validation sets according to a certain ratio (e.g., 7:3 or 8:2). The validation set is not learned at all during the entire training process and is only used to objectively evaluate the model performance.

[0069] Preprocessed image data from the training set is input into either the first or second model, which outputs a predicted pathological grade. The predicted result is compared with the actual pathological grade label, and the loss value is calculated. Backpropagation is employed, which accurately calculates the contribution of each parameter in the model to the final error. An optimizer (such as SGD or Adam) is used to iteratively update the model parameters based on the gradient information obtained from backpropagation, aiming to minimize the loss value.

[0070] After each training run on the training set, the system tests the current model on the validation set and records its classification accuracy. The system repeats the "train-validation" process and continuously tracks the validation set accuracy. Training terminates when the accuracy stops improving or even begins to decline (signs of overfitting), or when the validation set accuracy reaches a preset threshold (e.g., 90%). The system automatically selects the best-performing model parameters on the validation set, saves and packages them into a final, clinically applicable tumor magnetic resonance imaging analysis model.

[0071] In some embodiments, the step of inputting the preprocessed image data into the tumor magnetic resonance imaging analysis model to extract image evidence and predict the corresponding pathological grading results includes: extracting features from the preprocessed image data through the convolutional layer of the tumor magnetic resonance imaging analysis model to generate a feature map containing tumor edges, internal textures, and signal intensity; classifying the feature map through a fully connected layer; and outputting the corresponding pathological grading prediction results.

[0072] This example demonstrates how a trained model can automate the analysis and diagnosis of a new patient's MRI data. It details the model's internal workflow: from raw images to high-level features, and finally to classification decisions.

[0073] Deep feature extraction involves feeding preprocessed multi-sequence MRI data from new patients into the final tumor MRI image analysis model. The data flows through multiple convolutional layers of the model. Each layer performs convolution and activation operations, progressively extracting and combining features:

[0074] Shallow convolutional layers identify basic image features, such as tumor edges (sharp or blurred) and differences in signal intensity between different tissues. Deep convolutional layers combine basic features into more complex patterns, such as the internal texture of the tumor (uniform or non-uniform, presence of necrosis or cystic degeneration).

[0075] After a series of convolution and pooling operations, the model generates a series of highly abstract feature maps. These feature maps essentially encode the original image into a set of features most useful for the classification task. These feature maps are flattened and fed into fully connected layers for classification processing. The fully connected layers act like an expert committee, weighting and voting on these deep features.

[0076] Finally, the model outputs a probability distribution using functions such as Softmax, indicating the likelihood that the image belongs to each pathological grade (e.g., low risk, intermediate risk, high risk). The system ultimately selects the category with the highest probability as the pathological grade prediction result.

[0077] In some embodiments, the step of integrating visualization interpretation technology based on the pathological grading results to clearly present the key image regions that the model focuses on when making judgments includes: analyzing the decision-making process of the tumor magnetic resonance imaging analysis model using a gradient-weighted class activation mapping algorithm, generating a heat map reflecting the regions of interest of the model, overlaying the heat map onto the original magnetic resonance image, marking the key image regions of the tumor edge and internal necrotic areas or abnormal enhancement areas, and displaying them in a visual form.

[0078] This embodiment aims to solve the "black box" problem of AI models. By integrating advanced interpretable AI technology, it presents the model's decision-making basis to doctors in an intuitive, heatmap-like form, greatly enhancing the model's clinical credibility and practicality.

[0079] Decision process analysis involves the system invoking algorithms such as Gradient Weighted Class Activation Mapping (Grad-CAM) while the model makes predictions about new images.

[0080] The core principle of Grad-CAM is to use the gradient information flowing from the model to the final decision during prediction to infer the importance of each spatial location on the feature map to the final result.

[0081] The algorithm generates a weight distribution for the final convolutional layer feature maps based on gradient information. These weighted feature maps are combined and upsampled to generate a heatmap of the same size as the original image. In this heatmap, colors (from blue to red) represent the degree of influence of the region on the model's decision (from low to high). The system then overlays the generated heatmap onto the original T2-weighted magnetic resonance image in a semi-transparent manner.

[0082] In this way, radiologists can clearly see whether the model's "high-risk" assessment is based on the invasive features of the tumor's periphery in the image, the aggressive biological behavior indicated by the internal necrotic areas, or the high blood supply characteristics represented by the abnormally enhanced areas. This is not only a tool to verify the model's accuracy but also helps doctors discover subtle imaging features they might overlook.

[0083] In some embodiments, generating a tumor invasiveness assessment report based on the spatial distribution of the key image region and the corresponding key regions of tumor cell distribution and stromal infiltration in the postoperative pathological section includes: matching the spatial coordinates of the key image region with the spatial coordinates of the tumor cell distribution region and the stromal infiltration region marked by immunohistochemical staining in the postoperative pathological section, calculating the spatial overlap rate of the tumor cell distribution region and the stromal infiltration region, generating a quantitative assessment report on the extent and degree of tumor invasion based on the overlap rate, and attaching a key region comparison diagram.

[0084] This embodiment objectively and quantitatively correlates AI's imaging findings with the gold standard in pathology, a crucial step in elevating AI-assisted diagnosis to a "near-gold standard" level. It no longer relies solely on statistical performance but provides supporting biological evidence.

[0085] Multimodal data spatial matching acquires high-resolution digital images of postoperative pathological sections through the system, and marks areas of dense tumor cell distribution and stromal infiltration through immunohistochemical staining.

[0086] Through sophisticated image registration techniques, the spatial coordinate system of the MRI image is precisely aligned with the coordinate system of the digitized pathological slide. This typically requires first identifying anatomical landmarks (such as the urethra and seminal vesicles) on the pathological slide that are visible on the MRI.

[0087] The spatial coordinates of the obtained key image regions (highlighted areas on the heatmap) are correlated with the coordinates of the corresponding key regions on the pathological slides. The spatial overlap rate (e.g., Dice coefficient) between the highlighted areas on the MRI heatmap and the tumor / invasive areas on the pathological slides is calculated by computer. A higher overlap rate indicates more accurate localization by the model.

[0088] The system integrates pathological grading prediction, heatmaps, and spatial overlap to automatically generate a quantitative assessment report. Example report content: Diagnosis: Predicted pathological grade is high-risk (Gleason score ≥8). Key imaging evidence: The model highly focuses on the anterior margin of the lesion (compared to pathological sections, this area represents the site of tumor breach). Invasiveness assessment: "The model suggests a high probability of extracapsular invasion; after verification with postoperative pathology, the spatial overlap rate reaches 92%." The report will display the MRI image with overlaid heatmap and the corresponding pathological section image side-by-side, clearly marking key areas for comparison with arrows or outlines.

[0089] In some embodiments, the system further includes: a clinical data fusion module for acquiring the patient's clinical indicator data, including blood PSA value, patient age, and family history of tumors; the model acquisition module is further used to perform feature fusion of the clinical indicator data and the preprocessed image data, and train the tumor magnetic resonance imaging analysis model through a multiple learning algorithm, so that the tumor magnetic resonance imaging analysis model can combine imaging features and clinical features to predict pathological grading.

[0090] This embodiment significantly expands the system, upgrading it from a simple image analysis tool into a comprehensive clinical decision support system. It recognizes that prostate cancer diagnosis is multifactorial, and combining imaging features with key clinical indicators simulates the decision-making process of top-tier multidisciplinary expert teams, thereby achieving diagnostic accuracy beyond a single dimension of information.

[0091] Clinical data fusion module: responsible for obtaining patients' non-imaging data from the hospital information system.

[0092] Data types: Blood PSA level: the most important serum tumor marker for prostate cancer. Patient age: an important risk stratification factor. Family history of cancer: provides genetic background information. Feature fusion and multiple learning: In the later stages of the model, typically before fully connected layers, the above clinical indicator data (which need to be digitized and standardized first) are fused with deep features extracted from MRI images. This may be achieved by concatenating them into a longer feature vector. When the model module performs supervised learning as described in Example 5, the input data includes the fused features.

[0093] The system employs a multi-learning algorithm, meaning that during the same training process, the model simultaneously learns how to extract features from images and how to optimally combine and weight these image features with clinical features.

[0094] Through this end-to-end joint training, the final tumor MRI image analysis model can learn to weigh information. For example, when the PSA value is abnormally high, even if the imaging features are not very typical, the model may tend to give a higher risk score, which is more in line with clinical logic.

[0095] It should be noted that the acquisition of any information mentioned in the system is in accordance with relevant regulations and with the user's consent, and will not infringe on the user's privacy or violate relevant laws and regulations.

[0096] This application provides an embodiment of a non-invasive tumor invasiveness assessment method based on magnetic resonance imaging optimization. The execution device for the method is a computer device deployed in the non-invasive tumor invasiveness assessment system based on magnetic resonance imaging optimization provided in any embodiment of this application.

[0097] The provided method includes steps S101 to S103. The computer device can be a handheld terminal, a laptop computer, a wearable device, or a robot, etc. This is used to implement steps S101 to S103 and their corresponding embodiments.

[0098] It should be noted that the acquisition of any information mentioned in the provided methods is in compliance with relevant regulations and is carried out with the user's consent, and will not infringe on the user's privacy or violate relevant laws and regulations.

[0099] Step S101. Receive the user's conventional magnetic resonance imaging data, which includes T1-weighted image data, T2-weighted image data, diffusion-weighted imaging data, and dynamic contrast-enhanced image data;

[0100] Step S102. Preprocess the conventional magnetic resonance image data to obtain preprocessed image data; determine whether a disease-specific database with a large sample size exists. If it exists, obtain a base model pre-trained on natural images or a large number of other medical images, and input the preprocessed image data into the base model for fine-tuning to obtain a first model adapted to the tumor magnetic resonance image analysis task; if it does not exist, perform unsupervised pre-training using the characteristics of the preprocessed image data itself. By partially occluding the image and then having the base model predict, the base model learns to understand the internal structure of the medical image to obtain a second model.

[0101] Step S103. Obtain pathologically labeled sample data, input the pathologically labeled sample data into the first model or the second model for supervised learning to obtain a tumor magnetic resonance imaging analysis model; input the preprocessed image data into the tumor magnetic resonance imaging analysis model to extract image evidence and predict the corresponding pathological grading results; based on the pathological grading results, integrate visualization interpretation technology to clearly present the key image regions that the model focuses on when making judgments, the key image regions include tumor edges, internal necrosis areas, or abnormal enhancement areas; generate a tumor invasiveness assessment report based on the spatial distribution of the key image regions and the corresponding key regions of tumor cell distribution and stromal infiltration in the postoperative pathological sections.

[0102] In some embodiments, the preprocessing of the conventional magnetic resonance image data to obtain preprocessed image data includes: removing noise from the conventional magnetic resonance image data by performing an image denoising algorithm, normalizing the image grayscale values ​​using a standardization method, and aligning image data from different sequences to the same coordinate system based on a spatial registration algorithm to generate preprocessed image data.

[0103] In some embodiments, determining whether a disease-specific database with a large sample size exists includes: querying a preset database sample size threshold to determine whether the number of magnetic resonance imaging samples with pathological labels contained in the disease-specific database exceeds the threshold; if it exceeds the threshold, it is determined that a disease-specific database with a large sample size exists; otherwise, it is determined that it does not exist.

[0104] In some embodiments, the step of obtaining a pre-trained base model on natural images or a large number of other medical images, and inputting the preprocessed image data into the base model for fine-tuning to obtain a first model adapted to the tumor magnetic resonance image analysis task includes: obtaining a base model based on a convolutional neural network, retaining the convolutional layer parameters of the base model, replacing or adjusting the fully connected layer structure, inputting the preprocessed image data into the adjusted base model, and iteratively training the fully connected layer parameters through a transfer learning algorithm until the loss function value of the base model on the tumor magnetic resonance image feature extraction task converges to obtain the first model.

[0105] In some embodiments, if the aforementioned feature is not present, unsupervised pre-training is performed using the characteristics of the preprocessed image data itself. By partially occluding the image and then having the base model predict its features, the base model learns to understand the internal structure of the medical image, thus obtaining a second model. This includes: randomly selecting a local region in the preprocessed image data for occlusion, generating input data with an occlusion mask, inputting the input data into the base model, training the base model to predict the image features of the occluded region using a self-supervised learning algorithm, updating the model parameters based on the difference between the prediction results and the original image features, until the base model can accurately reconstruct the image structure of the unoccluded region, thus obtaining the second model.

[0106] In some embodiments, the step of inputting the pathologically labeled sample data into the first model or the second model for supervised learning to obtain a tumor magnetic resonance imaging analysis model includes: dividing the pathologically labeled sample data into a training set and a validation set; inputting the preprocessed image data and its corresponding pathological grading labels from the training set into the first model or the second model; optimizing the model parameters using a backpropagation algorithm; evaluating the accuracy of the model on the pathological grading task using a validation set; repeating the training and validation process until the accuracy of the first model or the second model on the validation set reaches a preset threshold, thereby obtaining the tumor magnetic resonance imaging analysis model.

[0107] In some embodiments, the step of inputting the preprocessed image data into the tumor magnetic resonance imaging analysis model to extract image evidence and predict the corresponding pathological grading results includes: extracting features from the preprocessed image data through the convolutional layer of the tumor magnetic resonance imaging analysis model to generate a feature map containing tumor edges, internal textures, and signal intensity; classifying the feature map through a fully connected layer; and outputting the corresponding pathological grading prediction results.

[0108] In some embodiments, the step of integrating visualization interpretation technology based on the pathological grading results to clearly present the key image regions that the model focuses on when making judgments includes: analyzing the decision-making process of the tumor magnetic resonance imaging analysis model using a gradient-weighted class activation mapping algorithm, generating a heat map reflecting the regions of interest of the model, overlaying the heat map onto the original magnetic resonance image, marking the key image regions of the tumor edge and internal necrotic areas or abnormal enhancement areas, and displaying them in a visual form.

[0109] In some embodiments, generating a tumor invasiveness assessment report based on the spatial distribution of the key image region and the corresponding key regions of tumor cell distribution and stromal infiltration in the postoperative pathological section includes: matching the spatial coordinates of the key image region with the spatial coordinates of the tumor cell distribution region and the stromal infiltration region marked by immunohistochemical staining in the postoperative pathological section, calculating the spatial overlap rate of the tumor cell distribution region and the stromal infiltration region, generating a quantitative assessment report on the extent and degree of tumor invasion based on the overlap rate, and attaching a key region comparison diagram.

[0110] In some embodiments, the system further includes: a clinical data fusion module for acquiring the patient's clinical indicator data, including blood PSA value, patient age, and family history of tumors; the model acquisition module is further used to perform feature fusion of the clinical indicator data and the preprocessed image data, and train the tumor magnetic resonance imaging analysis model through a multiple learning algorithm, so that the tumor magnetic resonance imaging analysis model can combine imaging features and clinical features to predict pathological grading.

[0111] It should be noted that those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the non-invasive tumor invasiveness assessment device and its modules based on magnetic resonance imaging optimization described above can be referred to the corresponding process in the embodiment of the non-invasive tumor invasiveness assessment system based on magnetic resonance imaging optimization described in any embodiment of this application, and will not be repeated here.

[0112] The provided non-invasive tumor invasiveness assessment method based on magnetic resonance imaging optimization can be implemented as a computer program that can run on the provided device.

[0113] The computer device provided in this application includes a processor, a memory, and a network interface connected via a device bus, wherein the memory may include a storage medium and internal memory.

[0114] The storage medium may store operating devices and computer programs. The computer program includes program instructions that, when executed, cause the processor to perform an embodiment of any non-invasive tumor invasiveness assessment method optimized based on magnetic resonance imaging.

[0115] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0116] Internal memory provides an environment for the execution of computer programs in non-volatile storage media. When executed by a processor, the computer program enables the processor to execute any non-invasive tumor invasiveness assessment system method based on magnetic resonance imaging optimization.

[0117] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that a specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0118] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0119] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps:

[0120] Receives conventional magnetic resonance imaging data from the user, including T1-weighted image data, T2-weighted image data, diffusion-weighted imaging data, and dynamic contrast-enhanced image data;

[0121] The conventional magnetic resonance imaging data is preprocessed to obtain preprocessed image data. It is then determined whether a disease-specific database with a large sample size exists. If it does, a pre-trained base model on natural images or a large number of other medical images is obtained. The preprocessed image data is input into the base model for fine-tuning to obtain a first model adapted to the tumor magnetic resonance imaging analysis task. If the database does not exist, unsupervised pre-training is performed using the characteristics of the preprocessed image data itself. By partially obscuring the image and then having the base model predict, the base model learns to understand the internal structure of the medical image, resulting in a second model.

[0122] Acquire pathologically labeled sample data, input the pathologically labeled sample data into the first model or the second model for supervised learning to obtain a tumor magnetic resonance imaging analysis model; input the preprocessed image data into the tumor magnetic resonance imaging analysis model to extract image evidence and predict the corresponding pathological grading results; based on the pathological grading results, integrate visualization interpretation technology to clearly present the key image regions that the model focuses on when making judgments, the key image regions including tumor margins, internal necrotic areas, or abnormal enhancement areas; generate a tumor invasiveness assessment report based on the spatial distribution of the key image regions and the corresponding key regions of tumor cell distribution and stromal infiltration in postoperative pathological sections.

[0123] In some embodiments, the preprocessing of the conventional magnetic resonance image data to obtain preprocessed image data includes: removing noise from the conventional magnetic resonance image data by performing an image denoising algorithm, normalizing the image grayscale values ​​using a standardization method, and aligning image data from different sequences to the same coordinate system based on a spatial registration algorithm to generate preprocessed image data.

[0124] In some embodiments, determining whether a disease-specific database with a large sample size exists includes: querying a preset database sample size threshold to determine whether the number of magnetic resonance imaging samples with pathological labels contained in the disease-specific database exceeds the threshold; if it exceeds the threshold, it is determined that a disease-specific database with a large sample size exists; otherwise, it is determined that it does not exist.

[0125] In some embodiments, the step of obtaining a pre-trained base model on natural images or a large number of other medical images, and inputting the preprocessed image data into the base model for fine-tuning to obtain a first model adapted to the tumor magnetic resonance image analysis task includes: obtaining a base model based on a convolutional neural network, retaining the convolutional layer parameters of the base model, replacing or adjusting the fully connected layer structure, inputting the preprocessed image data into the adjusted base model, and iteratively training the fully connected layer parameters through a transfer learning algorithm until the loss function value of the base model on the tumor magnetic resonance image feature extraction task converges to obtain the first model.

[0126] In some embodiments, if the aforementioned feature is not present, unsupervised pre-training is performed using the characteristics of the preprocessed image data itself. By partially occluding the image and then having the base model predict its features, the base model learns to understand the internal structure of the medical image, thus obtaining a second model. This includes: randomly selecting a local region in the preprocessed image data for occlusion, generating input data with an occlusion mask, inputting the input data into the base model, training the base model to predict the image features of the occluded region using a self-supervised learning algorithm, updating the model parameters based on the difference between the prediction results and the original image features, until the base model can accurately reconstruct the image structure of the unoccluded region, thus obtaining the second model.

[0127] In some embodiments, the step of inputting the pathologically labeled sample data into the first model or the second model for supervised learning to obtain a tumor magnetic resonance imaging analysis model includes: dividing the pathologically labeled sample data into a training set and a validation set; inputting the preprocessed image data and its corresponding pathological grading labels from the training set into the first model or the second model; optimizing the model parameters using a backpropagation algorithm; evaluating the accuracy of the model on the pathological grading task using a validation set; repeating the training and validation process until the accuracy of the first model or the second model on the validation set reaches a preset threshold, thereby obtaining the tumor magnetic resonance imaging analysis model.

[0128] In some embodiments, the step of inputting the preprocessed image data into the tumor magnetic resonance imaging analysis model to extract image evidence and predict the corresponding pathological grading results includes: extracting features from the preprocessed image data through the convolutional layer of the tumor magnetic resonance imaging analysis model to generate a feature map containing tumor edges, internal textures, and signal intensity; classifying the feature map through a fully connected layer; and outputting the corresponding pathological grading prediction results.

[0129] In some embodiments, the step of integrating visualization interpretation technology based on the pathological grading results to clearly present the key image regions that the model focuses on when making judgments includes: analyzing the decision-making process of the tumor magnetic resonance imaging analysis model using a gradient-weighted class activation mapping algorithm, generating a heat map reflecting the regions of interest of the model, overlaying the heat map onto the original magnetic resonance image, marking the key image regions of the tumor edge and internal necrotic areas or abnormal enhancement areas, and displaying them in a visual form.

[0130] In some embodiments, generating a tumor invasiveness assessment report based on the spatial distribution of the key image region and the corresponding key regions of tumor cell distribution and stromal infiltration in the postoperative pathological section includes: matching the spatial coordinates of the key image region with the spatial coordinates of the tumor cell distribution region and the stromal infiltration region marked by immunohistochemical staining in the postoperative pathological section, calculating the spatial overlap rate of the tumor cell distribution region and the stromal infiltration region, generating a quantitative assessment report on the extent and degree of tumor invasion based on the overlap rate, and attaching a key region comparison diagram.

[0131] In some embodiments, the system further includes: a clinical data fusion module for acquiring the patient's clinical indicator data, including blood PSA value, patient age, and family history of tumors; the model acquisition module is further used to perform feature fusion of the clinical indicator data and the preprocessed image data, and train the tumor magnetic resonance imaging analysis model through a multiple learning algorithm, so that the tumor magnetic resonance imaging analysis model can combine imaging features and clinical features to predict pathological grading.

[0132] It should be noted that those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the processor described above can be referred to the corresponding process in the method embodiments of the above embodiments, and will not be repeated here.

[0133] The embodiments of this application also provide a computer-readable storage medium storing a computer program, the computer program including program instructions, and the processor executing the program instructions to implement the steps of the non-invasive tumor invasiveness assessment method based on magnetic resonance imaging optimization provided in the above embodiments of this application.

[0134] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device.

[0135] It should be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. It should be understood that when an element or layer is referred to as “on,” “adjacent to,” “connected to,” or “coupled to” other elements or layers, it may be directly on, adjacent to, connected to, or coupled to other elements or layers, or there may be intervening elements or layers. Conversely, when an element is referred to as “directly on,” “directly adjacent to,” “directly connected to,” or “directly coupled to” other elements or layers, there are no intervening elements or layers. It should be understood that although the terms first, second, third, etc., may be used to describe various elements, components, areas, layers, and / or portions, these elements, components, areas, layers, and / or portions should not be limited by these terms. These terms are merely used to distinguish one element, component, area, layer, or portion from another element, component, area, layer, or portion. Therefore, without departing from the teachings of this application, the first element, component, area, layer, or portion discussed below may be referred to as a second element, component, area, layer, or portion.

[0136] Spatial relation terms such as “below,” “under,” “below,” “under,” “above,” “above,” etc., are used herein for convenience of description to describe the relationship between one element or feature shown in the figure and other elements or features. It should be understood that, in addition to the orientation shown in the figure, spatial relation terms are intended to also include different orientations of the device in use and operation. For example, if the device in the figure is flipped, then the element or feature described as “below,” “under,” or “below” other elements or features will be oriented “above” other elements or features. Therefore, the exemplary terms “below” and “under” can include both above and below orientations. The device may be otherwise oriented (rotated 90 degrees or otherwise) and the spatial descriptive terms used herein will be interpreted accordingly.

[0137] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. When used herein, the singular forms “a,” “an,” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising” and / or “including,” when used in this specification, identify the presence of the stated features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups. When used herein, the term “and / or” includes any and all combinations of the associated listed items.

[0138] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0139] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A non-invasive tumor invasiveness assessment system based on magnetic resonance imaging optimization, characterized in that, include: The data receiving module is used to receive the user's conventional magnetic resonance imaging data, which includes T1-weighted image data, T2-weighted image data, diffusion-weighted imaging data, and dynamic contrast-enhanced image data. The image processing module is used to preprocess the conventional magnetic resonance image data to obtain preprocessed image data; determine whether a disease-specific database with a large sample size exists; if it exists, obtain a base model pre-trained on natural images or a large number of other medical images, and input the preprocessed image data into the base model for fine-tuning to obtain a first model adapted to the tumor magnetic resonance image analysis task; if it does not exist, utilize the characteristics of the preprocessed image data itself for unsupervised pre-training, and by partially occluding the image and then having the base model predict, enable the base model to learn to understand the internal structure of the medical image to obtain a second model; The model acquisition module is used to acquire sample data with pathological labels, input the sample data with pathological labels into the first model or the second model for supervised learning, and obtain a tumor magnetic resonance imaging analysis model; input the preprocessed image data into the tumor magnetic resonance imaging analysis model, extract image evidence, and predict the corresponding pathological grading results; The report generation module is used to integrate visualization interpretation technology based on the pathological grading results to clearly present the key image regions that the model focuses on when making judgments. The key image regions include tumor edges, internal necrotic areas, or abnormal enhancement areas. Based on the spatial distribution of the key image regions and the corresponding key regions of tumor cell distribution and stromal infiltration in the postoperative pathological sections, a tumor invasiveness assessment report is generated.

2. The system according to claim 1, characterized in that, The preprocessing of the conventional magnetic resonance image data to obtain preprocessed image data includes: The conventional magnetic resonance image data is noise-removed by performing an image denoising algorithm, the image grayscale values ​​are normalized by a standardization method, and the image data of different sequences are aligned to the same coordinate system based on a spatial registration algorithm to generate preprocessed image data.

3. The system according to claim 1, characterized in that, The determination of whether a disease-specific database with a large sample size exists includes: By querying a preset database sample size threshold, it is determined whether the number of magnetic resonance image samples with pathological labels contained in the disease-specific database exceeds the threshold. If it exceeds the threshold, it is determined that there is a disease-specific database with a large sample size; otherwise, it is determined that there is no such database.

4. The system according to claim 1, characterized in that, The process of acquiring a pre-trained base model on natural images or a large number of other medical images, and inputting the preprocessed image data into the base model for fine-tuning to obtain a first model adapted to the tumor magnetic resonance image analysis task includes: A base model based on a convolutional neural network is obtained. The convolutional layer parameters of the base model are retained, and the fully connected layer structure is replaced or adjusted. The preprocessed image data is input into the adjusted base model, and the fully connected layer parameters are iteratively trained through a transfer learning algorithm until the loss function value of the base model on the tumor magnetic resonance image feature extraction task converges, thus obtaining the first model.

5. The system according to claim 1, characterized in that, If the aforementioned features are not present, unsupervised pre-training is performed using the characteristics of the preprocessed image data itself. By partially occluding the image and then having the base model predict, the base model learns to understand the internal structure of the medical image, resulting in a second model, which includes: A local region in the preprocessed image data is randomly selected for occlusion, generating input data with an occlusion mask. The input data is then input into a base model, which is trained using a self-supervised learning algorithm to predict the image features of the occluded region. The model parameters are updated based on the difference between the prediction results and the original image features until the base model can accurately reconstruct the image structure of the unoccluded region, thus obtaining the second model.

6. The system according to claim 1, characterized in that, The step of inputting the pathologically labeled sample data into the first model or the second model for supervised learning to obtain a tumor magnetic resonance imaging analysis model includes: The pathologically labeled sample data is divided into a training set and a validation set. The preprocessed image data and their corresponding pathological grading labels in the training set are input into the first model or the second model. The backpropagation algorithm is used to optimize the model parameters. The accuracy of the model on the pathological grading task is evaluated through the validation set. The training and validation process is repeated until the accuracy of the first model or the second model on the validation set reaches a preset threshold, thus obtaining the tumor magnetic resonance imaging analysis model.

7. The system according to claim 1, characterized in that, The step of inputting the preprocessed image data into the tumor magnetic resonance imaging analysis model to extract image evidence and predict the corresponding pathological grading results includes: The preprocessed image data is feature extracted by the convolutional layer of the tumor magnetic resonance imaging analysis model to generate a feature map containing tumor edges, internal texture and signal intensity. The feature map is then classified by a fully connected layer to output the corresponding pathological grade prediction result.

8. The system according to claim 1, characterized in that, Based on the pathological grading results, the integrated visualization interpretation technology clearly presents the key image regions that the model focuses on when making judgments, including: The decision-making process of the tumor magnetic resonance imaging analysis model is analyzed using a gradient-weighted class activation mapping algorithm to generate a heat map reflecting the model's areas of interest. The heat map is then overlaid onto the original magnetic resonance image to mark key image regions such as the tumor margin and internal necrotic or abnormally enhanced areas, and presented in a visual format.

9. The system according to claim 1, characterized in that, The process involves generating a tumor invasiveness assessment report based on the spatial distribution of the key image regions and the corresponding key regions of tumor cell distribution and stromal infiltration in the postoperative pathological sections. This report includes: The spatial coordinates of the key image region are matched with the spatial coordinates of the tumor cell distribution area and the stromal infiltration area marked by immunohistochemical staining in the postoperative pathological section. The spatial overlap rate of the tumor cell distribution area and the stromal infiltration area is calculated. Based on the overlap rate, a quantitative assessment report on the extent and degree of tumor invasion is generated, along with a comparison diagram of the key areas.

10. The system according to claim 1, characterized in that, The system also includes: The clinical data fusion module is used to acquire the patient's clinical indicator data, including blood PSA value, patient age, and family history of tumors. The model acquisition module is also used to fuse the clinical indicator data with the preprocessed image data, and train the tumor magnetic resonance imaging analysis model through multiple learning algorithms, so that the tumor magnetic resonance imaging analysis model can combine imaging features and clinical features to predict pathological grading.

Images