A Deep Learning-Driven 3D Medical Imaging Diagnostic Assistance System

Abstract

This invention discloses a deep learning-driven 3D medical image diagnostic assistance system. In this system, the model training module serves as the core, and the meticulous design and collaborative work of its internal sub-modules bring significant benefits to the entire system. The data loading sub-module ensures efficient reading and preprocessing of large-scale image data, providing a stable data flow for subsequent training. Through automated feature learning and model customization and optimization, the model training module achieves efficient transformation from raw image data to diagnostic knowledge, greatly improving diagnostic accuracy and efficiency. Doctors can rely on the accurate diagnostic results provided by the system to develop treatment plans more quickly and improve patient prognosis. Simultaneously, efficient training management and performance monitoring and evaluation ensure the stability and reliability of the model throughout the training process.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional medical imaging technology, specifically a deep learning-driven three-dimensional medical imaging diagnostic auxiliary system. Background Technology

[0002] Three-dimensional medical imaging diagnosis primarily utilizes equipment such as computed tomography (CT), magnetic resonance imaging (MRI), and ultrasound imaging to acquire three-dimensional images of tissues within the patient's body. With the continuous advancement of medical technology, three-dimensional medical imaging, including MRI, CT, and PET, plays an increasingly important role in clinical diagnosis. However, traditional medical imaging diagnosis relies heavily on the experience and visual judgment of physicians, resulting in problems such as high subjectivity, low efficiency, and a high risk of missed diagnoses. To improve diagnostic accuracy and efficiency, deep learning technology has been widely applied in the field of medical image analysis in recent years. Three-dimensional medical imaging diagnosis, as an advanced medical diagnostic technology, is widely used in clinical diagnosis, disease treatment, and medical research.

[0003] However, existing technologies suffer from inaccurate feature extraction, rely on manually designed features, and struggle to fully capture the deep information in complex 3D images; model training is difficult, and they are prone to overfitting or underfitting. Furthermore, existing diagnostic predictions are inflexible and fail to meet the diagnostic needs of various diseases; the results are not presented intuitively, lacking both visual appeal and interactivity. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-driven three-dimensional medical image diagnostic assistance system in order to solve the problems mentioned above.

[0005] The technical solution adopted in this invention is as follows: a deep learning-driven three-dimensional medical image diagnostic auxiliary system, the system comprising: an image data acquisition module, a data preprocessing module, a feature extraction module, a model training module, a diagnostic prediction module, and a result display module;

[0006] The model training module is internally configured with a data preparation submodule, a model design submodule, a training process management submodule, a performance evaluation submodule, a hyperparameter tuning submodule, and a model saving and loading submodule;

[0007] The image data acquisition module is responsible for acquiring raw 3D image data and corresponding annotation information from medical imaging equipment. This data is then transmitted to the data preprocessing module. The data preprocessing module performs denoising, alignment, correction, normalization, cropping or scaling operations on the raw image data, converts it into a format that meets the model input requirements, and stores it for subsequent use. The preprocessed data then enters the feature extraction module.

[0008] The feature extraction module automatically extracts key features from the image using three-dimensional convolutional neural network deep learning technology. These features will be used to describe the structure and patterns in the image.

[0009] The model training module trains the model using features and corresponding labeled data, and adjusts the model parameters through optimization algorithms to minimize prediction errors. The trained model is then used in the diagnostic prediction module to predict new image data and generate diagnostic results.

[0010] The results display module presents the model's prediction results to doctors and patients in an intuitive way, including lesion annotations on 3D images, predicted probabilities, and detailed diagnostic reports. It also displays the changes in performance indicators during model training, thus completing the entire process from data collection to results display.

[0011] In a preferred embodiment, the image data acquisition module is responsible for acquiring raw three-dimensional image data from various medical imaging devices; scanning the patient's brain using a magnetic resonance imaging device to generate a series of two-dimensional slices, which are then synthesized into a three-dimensional image; during the data acquisition process, it is necessary to ensure that the resolution, contrast, and signal-to-noise ratio of the image meet the requirements of subsequent processing; in addition, it is also necessary to collect patient information corresponding to the image, including age, gender, and clinical diagnosis, which will be used as annotation data along with the image data for subsequent analysis and model training; after the data acquisition is completed, the image data and annotation data are stored in a secure database for subsequent access and processing.

[0012] In a preferred embodiment, the data preprocessing module first performs denoising on the 3D images to reduce noise introduced by device or environmental factors; then, it performs image alignment and correction to ensure spatial consistency of images from different patients; furthermore, it performs image normalization to scale pixel values ​​to a uniform range, such as 0 to 1, to eliminate differences in brightness and contrast between different images; preprocessing also includes image cropping or scaling to match the size requirements of the model input; finally, as needed, it performs data augmentation operations, including random rotation, flipping, or scaling, to increase the generalization ability of the model.

[0013] In a preferred embodiment, the data preparation submodule specifically includes:

[0014] Data loading: Load the preprocessed 3D image data X and the corresponding annotation data Y;

[0015] X: A tensor of shape (N,C,D,H,W), where N is the number of samples, C is the number of channels, and D, H, and W are the depth, height, and width, respectively.

[0016] Y: A tensor of shape (N,L), where L is the number of labels;

[0017] Data augmentation: Perform random rotation, scaling, and flipping operations on X to generate augmented data X_aug;

[0018] Data splitting: Split X_aug and Y into training set (X_train, Y_train), validation set (X_val, Y_val), and test set (X_test, Y_test).

[0019] In a preferred embodiment, the model design submodule specifically includes:

[0020] Network architecture selection: 3DCNN was chosen as the basic architecture;

[0021] Layer and parameter configuration:

[0022] Input layer: Accepts inputs with shapes (C,D,H,W);

[0023] Convolutional layers: Multiple 3D convolutional layers are used, and the parameters of each convolutional layer include kernel size k, stride s, and padding p;

[0024] Activation function: Use the ReLU function;

[0025] Pooling layer: Uses 3D max pooling layer, with parameters including pooling kernel size k_pool and stride s_pool;

[0026] Fully connected layer: The feature map is flattened and then connected to the fully connected layer, and finally the probabilities of L classes are output;

[0027] Optimizer selection: Select the Adam optimizer, with parameters including learning rate α and weight decay λ.

[0028] In a preferred embodiment, the training process management submodule specifically includes:

[0029] Batch processing: Process X_train and Y_train in batches, with each batch being B in size;

[0030] Forward propagation: For each batch of data X_batch, the output Y_pred is calculated using a 3DCNN;

[0031] Y_pred = CNN(X_batch)

[0032] Loss Calculation

[0033] Calculate the loss L using the cross-entropy loss function;

[0034] L = -Σ(Y_true*log(Y_pred)) / B

[0035] Where Y_true is the true label and Y_pred is the probability distribution predicted by the model;

[0036] Backpropagation

[0037] Calculate the gradient of the loss function with respect to the network parameters;

[0038] Parameter update

[0039] Update network parameters using the Adam optimizer;

[0040]

[0041] Where θ is the network parameter and α is the learning rate. It is the gradient of the loss function with respect to θ.

[0042] In a preferred embodiment, the performance evaluation submodule specifically includes:

[0043] Metrics calculation: Calculate accuracy (Acc), recall (Rec), and F1 score (F1).

[0044] Acc = (TP + TN) / (P + N)

[0045] Rec = TP / P

[0046] F1 = 2*(Prec*Rec) / (Prec+Rec)

[0047] Where TP represents true positives, TN represents true negatives, P represents the total number of positives, N represents the total number of negatives, and Prec represents the precision.

[0048] Validation set testing: Test the model performance on the validation set and calculate the above metrics;

[0049] Model selection: Select the best model based on performance metrics on the validation set;

[0050] The hyperparameter tuning submodule specifically includes:

[0051] Hyperparameter search: Use grid search or random search methods to find the optimal combination of hyperparameters, including learning rate α and batch size B;

[0052] Cross-validation: K-fold cross-validation is used to evaluate model performance under different combinations of hyperparameters;

[0053] The model saving and loading submodule saves the trained model parameters to a file for later use.

[0054] In a preferred embodiment, the feature extraction module preprocesses and extracts features from the acquired raw image data. First, the 3D image undergoes denoising, alignment, and normalization preprocessing to reduce the impact of noise and variability. Then, a specific algorithm is used to extract useful features from the preprocessed image. These features include morphological features, texture features, and intensity features. The purpose of feature extraction is to convert high-dimensional image data into lower-dimensional feature representations that are easier for the model to process. The extracted features will be stored and used for subsequent model training.

[0055] In a preferred embodiment, the diagnostic prediction module uses a trained model to predict the diagnosis of new image data. When new patient image data is input, the images are first subjected to the same preprocessing and feature extraction operations, and then the extracted features are input into the trained 3DCNN model. The model calculates the output through forward propagation to obtain the probability that the patient belongs to different diagnostic categories. Based on these probabilities, the optimal diagnostic result for the patient is determined. If the model outputs that the probability of the patient having a certain disease exceeds a set threshold, then the patient is diagnosed with that disease. The diagnostic prediction results are recorded and used for subsequent clinical decision-making.

[0056] In a preferred embodiment, the results display module presents the model's diagnostic prediction results to doctors and patients in an intuitive and easy-to-understand manner; it displays the patient's three-dimensional image using a graphical user interface and marks the lesion areas predicted by the model on the image; it also provides a detailed diagnostic report, including the model's predicted probability, disease type, and suggested further examinations and treatment plans; furthermore, it uses visualization tools to display changes in loss and performance indicators during model training, so that doctors and researchers can evaluate the model's performance and stability; the goal of the results display module is to provide comprehensive and accurate information to assist doctors in making better clinical decisions.

[0057] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0058] 1. In this invention, the model training module, as the core of the deep learning-driven 3D medical image diagnostic assistance system, brings significant benefits to the entire system through the meticulous design and collaborative work of its internal sub-modules. The data loading sub-module ensures efficient reading and preprocessing of large-scale image data, providing a stable data flow for subsequent training. Through automated feature learning and model customization and optimization, the model training module enables the system to efficiently transform raw image data into diagnostic knowledge, greatly improving the accuracy and efficiency of diagnosis. Doctors can rely on the accurate diagnostic results provided by the system to formulate treatment plans more quickly and improve patient prognosis. Simultaneously, efficient training management and performance monitoring and evaluation ensure the stability and reliability of the model throughout the training process, enabling the finally deployed model to perform excellently in real clinical environments, providing doctors with continuous and accurate auxiliary diagnostic services. Furthermore, the implementation of model reuse and deployment functions allows the trained model to be easily saved, shared, and deployed to different medical environments, further expanding the system's application scope and influence.

[0059] 2. In this invention, the image data acquisition module ensures the accurate acquisition of raw data, providing high-quality basic materials for subsequent analysis and guaranteeing the reliability and effectiveness of the diagnosis. The data preprocessing module improves data consistency and usability through operations such as denoising, alignment, correction, and normalization, reducing diagnostic errors caused by data quality issues and laying a solid foundation for feature extraction and model training. The feature extraction module automatically extracts key features using deep learning technology, greatly improving the efficiency and accuracy of feature extraction, enabling the model to better capture lesion information in the images, thereby improving the accuracy of the diagnosis.

[0060] 3. In this invention, the diagnostic prediction module utilizes a trained model to quickly and accurately predict new image data, providing doctors with strong decision support, shortening diagnostic time, and improving diagnostic efficiency. The results display module presents the diagnostic results in an intuitive and easy-to-understand manner, including lesion annotations on 3D images, predicted probabilities, and detailed diagnostic reports, enabling doctors and patients to clearly understand the diagnostic situation, enhancing doctor-patient communication, and increasing patient trust and satisfaction. The organic combination of these modules forms a highly efficient, accurate, and easy-to-use 3D medical imaging diagnostic assistance system, which not only improves the efficiency and quality of medical diagnosis but also promotes the rapid development of medical imaging diagnosis towards intelligence and precision, bringing revolutionary progress to the medical industry. Attached Figure Description

[0061] Figure 1 This is an overall system block diagram of the present invention;

[0062] Figure 2 This is a system block diagram of the model training module in this invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0064] Reference Figure 1-2 ,

[0065] A deep learning-driven 3D medical image diagnostic assistance system, comprising: an image data acquisition module, a data preprocessing module, a feature extraction module, a model training module, a diagnostic prediction module, and a result display module;

[0066] The model training module is internally configured with a data preparation submodule, a model design submodule, a training process management submodule, a performance evaluation submodule, a hyperparameter tuning submodule, and a model saving and loading submodule;

[0067] The image data acquisition module is responsible for acquiring raw 3D image data and corresponding annotation information from medical imaging equipment. This data is then transmitted to the data preprocessing module. The data preprocessing module performs operations such as denoising, alignment, correction, normalization, and possible cropping or scaling on the raw image data, converting it into a format that meets the model input requirements and storing it for subsequent use. The preprocessed data then enters the feature extraction module.

[0068] The feature extraction module automatically extracts key features from images using deep learning techniques such as 3D convolutional neural networks (3D CNN). These features will be used to describe the structure and patterns in the images.

[0069] The model training module trains the model using features and corresponding labeled data, and adjusts model parameters through optimization algorithms to minimize prediction error. The trained model is then used in the diagnostic prediction module to predict new image data and generate diagnostic results.

[0070] The results display module presents the model's predictions to doctors and patients in an intuitive way, including lesion annotations on 3D images, predicted probabilities, and detailed diagnostic reports. It can also display changes in performance metrics during model training, thus completing the entire process from data collection to results display.

[0071] The image data acquisition module is responsible for acquiring raw 3D image data from various medical imaging devices. Magnetic resonance imaging (MRI) scans the patient's brain, generating a series of 2D slices, which are then synthesized into a 3D image. During data acquisition, it is crucial to ensure that the image resolution, contrast, and signal-to-noise ratio meet the requirements of subsequent processing. Furthermore, patient information corresponding to the images, such as age, gender, and clinical diagnosis, needs to be collected. This information will be used as annotation data along with the image data for subsequent analysis and model training. After data acquisition, the image data and annotation data are stored in a secure database for easy access and processing.

[0072] In the data preprocessing module, a series of preprocessing operations are performed on the acquired raw image data to ensure data quality and consistency, and to prepare for subsequent feature extraction and model training. First, denoising is performed on the 3D images to reduce noise introduced by equipment or environmental factors. Next, image alignment and correction are performed to ensure spatial consistency between images of different patients. Furthermore, image normalization may be performed, scaling pixel values ​​to a uniform range, such as 0 to 1, to eliminate differences in brightness and contrast between different images. Preprocessing may also include image cropping or scaling to match the size requirements of the model input. Finally, data augmentation operations, such as random rotation, flipping, or scaling, may be performed as needed to increase the model's generalization ability. The preprocessed image data will be stored and used for subsequent feature extraction and model training.

[0073] The data preparation submodule specifically includes:

[0074] Data loading:

[0075] Load the preprocessed 3D image data X and the corresponding annotation data Y.

[0076] X: A tensor of shape (N,C,D,H,W), where N is the number of samples, C is the number of channels, and D, H, and W are the depth, height, and width, respectively.

[0077] Y: A tensor of shape (N,L), where L is the number of labels.

[0078] Data augmentation:

[0079] Perform random rotation, scaling, and flipping operations on X to generate augmented data X_aug.

[0080] Data partitioning:

[0081] Split X_aug and Y into a training set (X_train, Y_train), a validation set (X_val, Y_val), and a test set (X_test, Y_test).

[0082] The model design submodule specifically includes:

[0083] Network architecture selection:

[0084] 3DCNN was chosen as the basic architecture.

[0085] Layer and parameter configuration:

[0086] Input layer: Accepts inputs with shape (C,D,H,W).

[0087] Convolutional layers: Multiple 3D convolutional layers are used, and the parameters of each convolutional layer include kernel size k, stride s, and padding p.

[0088] Activation function: The ReLU function is usually used.

[0089] Pooling layer: Uses 3D max pooling layer, with parameters including pooling kernel size k_pool and stride s_pool.

[0090] Fully connected layer: The feature map is flattened and then connected to the fully connected layer, and finally the probabilities of L categories are output.

[0091] Optimizer selection:

[0092] Choose the Adam optimizer, with parameters including learning rate α, weight decay λ, etc.

[0093] The training process management submodule specifically includes:

[0094] Batch processing: Process X_train and Y_train in batches, with each batch being B in size.

[0095] Forward propagation: For each batch of data X_batch, the output Y_pred is calculated using a 3DCNN.

[0096] Y_pred = CNN(X_batch)

[0097] Loss Calculation

[0098] The loss L is calculated using the cross-entropy loss function.

[0099] L = -Σ(Y_true*log(Y_pred)) / B

[0100] Where Y_true is the true label and Y_pred is the probability distribution predicted by the model.

[0101] Backpropagation

[0102] Calculate the gradient of the loss function with respect to the network parameters.

[0103] Parameter update

[0104] Update network parameters using the Adam optimizer.

[0105]

[0106] Where θ is the network parameter and α is the learning rate. It is the gradient of the loss function with respect to θ.

[0107] The performance evaluation submodule specifically includes:

[0108] Metrics calculation: Calculate accuracy (Acc), recall (Rec), and F1 score (F1).

[0109] Acc = (TP + TN) / (P + N)

[0110] Rec = TP / P

[0111] F1 = 2*(Prec*Rec) / (Prec+Rec)

[0112] Where TP represents true positives, TN represents true negatives, P represents the total number of positives, N represents the total number of negatives, and Prec represents the precision.

[0113] Validation set testing: Test the model performance on the validation set and calculate the above metrics.

[0114] Model selection: Select the best model based on performance metrics on the validation set;

[0115] The hyperparameter tuning submodule specifically includes:

[0116] Hyperparameter search: Use grid search or random search methods to find the optimal combination of hyperparameters, such as learning rate α and batch size B.

[0117] Cross-validation: K-fold cross-validation is used to evaluate model performance under different combinations of hyperparameters;

[0118] The model saving and loading submodule saves the trained model parameters to a file for later use.

[0119] The feature extraction module preprocesses and extracts features from the acquired raw image data. First, the 3D image undergoes preprocessing operations such as denoising, alignment, and normalization to reduce the impact of noise and variability. Then, a specific algorithm extracts useful features from the preprocessed image. These features may include morphological features (such as volume and shape), texture features (such as the gray-level co-occurrence matrix), and intensity features (such as average intensity and standard deviation). The purpose of feature extraction is to transform high-dimensional image data into a lower-dimensional feature representation that is easier for the model to process. The extracted features are stored and used for subsequent model training.

[0120] In the diagnostic prediction module, a trained model is used to predict diagnoses based on new image data. When new patient image data is input, the images undergo the same preprocessing and feature extraction operations. The extracted features are then fed into the trained 3D CNN model. The model calculates the probability of the patient belonging to different diagnostic categories through forward propagation. Based on these probabilities, the most likely diagnosis can be determined. If the model outputs a probability that the patient has a certain disease exceeding a set threshold (e.g., 0.5), that disease can be diagnosed. The diagnostic prediction results are recorded and used for subsequent clinical decision-making.

[0121] The results presentation module displays the model's diagnostic predictions in an intuitive and easy-to-understand manner to doctors and patients. A graphical user interface (GUI) can be used to display the patient's 3D images, with the predicted lesion areas annotated on the images. Detailed diagnostic reports are also available, including the model's predicted probabilities, possible disease types, and suggested further examinations and treatments. Furthermore, visualization tools (such as TensorBoard) can be used to show changes in loss and performance metrics during model training, facilitating evaluation of the model's performance and stability by doctors and researchers. The goal of the results presentation module is to provide comprehensive and accurate information to assist doctors in making better clinical decisions.

[0122] In this invention, the model training module, as the core of the deep learning-driven 3D medical image diagnostic assistance system, brings significant benefits to the entire system through the meticulous design and collaborative work of its internal sub-modules. The data loading sub-module ensures efficient reading and preprocessing of large-scale image data, providing a stable data flow for subsequent training. Through automated feature learning and model customization and optimization, the model training module enables the system to efficiently transform raw image data into diagnostic knowledge, greatly improving diagnostic accuracy and efficiency. Doctors can rely on the accurate diagnostic results provided by the system to develop treatment plans more quickly and improve patient prognosis. Simultaneously, efficient training management and performance monitoring and evaluation ensure the stability and reliability of the model throughout the training process, enabling the finally deployed model to perform excellently in real clinical environments, providing doctors with continuous and accurate auxiliary diagnostic services. Furthermore, the implementation of model reuse and deployment functions allows the trained model to be easily saved, shared, and deployed to different medical environments, further expanding the system's application scope and influence.

[0123] In this invention, the image data acquisition module ensures the accurate acquisition of raw data, providing high-quality foundational material for subsequent analysis and guaranteeing the reliability and effectiveness of the diagnosis. The data preprocessing module, through operations such as denoising, alignment, correction, and normalization, improves data consistency and usability, reduces diagnostic errors caused by data quality issues, and lays a solid foundation for feature extraction and model training. The feature extraction module utilizes deep learning technology to automatically extract key features, greatly improving the efficiency and accuracy of feature extraction, enabling the model to better capture lesion information in the images, thereby enhancing diagnostic precision.

[0124] In this invention, the diagnostic prediction module utilizes a trained model to quickly and accurately predict new image data, providing doctors with strong decision support, shortening diagnostic time, and improving diagnostic efficiency. The results display module presents the diagnostic results in an intuitive and easy-to-understand manner, including lesion annotations on 3D images, predicted probabilities, and detailed diagnostic reports, enabling doctors and patients to clearly understand the diagnosis, enhancing doctor-patient communication, and increasing patient trust and satisfaction. The organic combination of these modules forms a highly efficient, accurate, and easy-to-use 3D medical imaging diagnostic assistance system, which not only improves the efficiency and quality of medical diagnosis but also promotes the rapid development of medical imaging diagnosis towards intelligence and precision, bringing revolutionary progress to the medical industry.

[0125] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0126] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-driven three-dimensional medical image diagnostic assistance system, characterized in that: The system includes: an image data acquisition module, a data preprocessing module, a feature extraction module, a model training module, a diagnosis and prediction module, and a result display module; The model training module is internally configured with a data preparation submodule, a model design submodule, a training process management submodule, a performance evaluation submodule, a hyperparameter tuning submodule, and a model saving and loading submodule; The image data acquisition module is responsible for acquiring raw 3D image data and corresponding annotation information from medical imaging equipment. This data is then transmitted to the data preprocessing module. The data preprocessing module performs denoising, alignment, correction, normalization, cropping or scaling operations on the raw image data, converts it into a format that meets the model input requirements, and stores it for subsequent use. The preprocessed data then enters the feature extraction module. The feature extraction module automatically extracts key features from the image using three-dimensional convolutional neural network deep learning technology. These features will be used to describe the structure and patterns in the image. The model training module trains the model using features and corresponding labeled data, and adjusts the model parameters through optimization algorithms to minimize prediction errors. The trained model is then used in the diagnostic prediction module to predict new image data and generate diagnostic results. The results display module presents the model's predictions to doctors and patients in an intuitive way, including lesion annotations on 3D images, predicted probabilities, and detailed diagnostic reports. It also displays performance metrics changes during model training, thus completing the entire process from data acquisition to results display. The image data acquisition module is responsible for acquiring raw 3D image data from various medical imaging devices; using magnetic resonance imaging (MRI) to scan the patient's brain, generating a series of 2D slices, which are then synthesized into 3D images; during data acquisition, it is necessary to ensure that the image resolution, contrast, and signal-to-noise ratio meet the requirements of subsequent processing; in addition, it is necessary to collect patient information corresponding to the images, including age, gender, and clinical diagnosis. This information will be used as annotation data along with the image data for subsequent analysis and model training; after data acquisition, the image data and annotation data are stored in a secure database for subsequent access and processing. The data preparation submodule specifically includes: Data loading: Load the preprocessed 3D image data X and the corresponding annotation data Y; X: A tensor of shape (N,C,D,H,W), where N is the number of samples, C is the number of channels, and D, H, and W are the depth, height, and width, respectively. Y: A tensor of shape (N,L), where L is the number of labels; Data augmentation: Perform random rotation, scaling, and flipping operations on X to generate augmented data X_aug; Data splitting: Split X_aug and Y into training set (X_train, Y_train), validation set (X_val, Y_val), and test set (X_test, Y_test).

2. The deep learning-driven three-dimensional medical image diagnostic assistance system as described in claim 1, characterized in that: The data preprocessing module first denoises the 3D images to reduce noise introduced by equipment or environmental factors. Next, it aligns and corrects the images to ensure spatial consistency across different patients. Furthermore, it normalizes the images, scaling pixel values ​​to a uniform range, such as 0 to 1, to eliminate differences in brightness and contrast between images. Preprocessing also includes image cropping or scaling to match the size requirements of the model input. Finally, data augmentation operations are performed as needed, including random rotation, flipping, or scaling, to increase the model's generalization ability.

3. The deep learning-driven three-dimensional medical image diagnostic assistance system as described in claim 1, characterized in that: The model design submodule specifically includes: Network architecture selection: 3DCNN was chosen as the basic architecture; Layer and parameter configuration: Input layer: Accepts inputs with shapes (C,D,H,W); Convolutional layers: Multiple 3D convolutional layers are used, and the parameters of each convolutional layer include kernel size k, stride s, and padding p; Activation function: Use the ReLU function; Pooling layer: Uses 3D max pooling layer, with parameters including pooling kernel size k_pool and stride s_pool; Fully connected layer: The feature map is flattened and then connected to the fully connected layer, and finally the probabilities of L classes are output; Optimizer selection: Select the Adam optimizer, with parameters including learning rate α and weight decay λ.

4. The deep learning-driven three-dimensional medical image diagnostic assistance system as described in claim 1, characterized in that: The training process management submodule specifically includes: Batch processing: Process X_train and Y_train in batches, with each batch being B in size; Forward propagation: For each batch of data X_batch, the output Y_pred is calculated using a 3DCNN; Y_pred=CNN(X_batch) Loss Calculation Calculate the loss L using the cross-entropy loss function; L=-Σ(Y_true*log(Y_pred)) / B Where Y_true is the true label and Y_pred is the probability distribution predicted by the model; Backpropagation Calculate the gradient of the loss function with respect to the network parameters; Parameter update Update network parameters using the Adam optimizer; θ=θ-α*∇_θL Where θ is the network parameter, α is the learning rate, and ∇_θL is the gradient of the loss function with respect to θ.

5. The deep learning-driven three-dimensional medical image diagnostic assistance system as described in claim 1, characterized in that: The performance evaluation submodule specifically includes: Metrics calculation: Calculate accuracy (Acc), recall (Rec), and F1 score (F1). Acc = (TP + TN) / (P + N) Rec=TP / P F1 = 2 * (Prec * Rec) / (Prec + Rec) Where TP represents true positives, TN represents true negatives, P represents the total number of positives, N represents the total number of negatives, and Prec represents the precision. Validation set testing: Test the model performance on the validation set and calculate the above metrics; Model selection: Select the best model based on performance metrics on the validation set; The hyperparameter tuning submodule specifically includes: Hyperparameter search: Use grid search or random search methods to find the optimal combination of hyperparameters, including learning rate α and batch size B; Cross-validation: K-fold cross-validation is used to evaluate model performance under different combinations of hyperparameters; The model saving and loading submodule saves the trained model parameters to a file for later use.

6. The deep learning-driven three-dimensional medical image diagnostic assistance system as described in claim 1, characterized in that: In the feature extraction module, the acquired raw image data is preprocessed and features are extracted. First, the 3D image is preprocessed by denoising, alignment and normalization to reduce the impact of noise and variability. Then, a specific algorithm is used to extract useful features from the preprocessed image. These features include morphological features, textural features, and intensity features; The purpose of feature extraction is to transform high-dimensional image data into lower-dimensional feature representations that are easier for models to process; the extracted features will be stored and used for subsequent model training.

7. The deep learning-driven three-dimensional medical image diagnostic assistance system as described in claim 1, characterized in that: In the diagnostic prediction module, a trained model is used to predict diagnoses based on new image data. When new patient image data is input, the images are first subjected to the same preprocessing and feature extraction operations, and then the extracted features are input into the trained 3DCNN model. The model calculates the probability of the patient belonging to different diagnostic categories through forward propagation. Based on these probabilities, the optimal diagnostic result for the patient is determined. If the model outputs that the probability of the patient having a certain disease exceeds a set threshold, then the patient is diagnosed with that disease. The diagnostic prediction results are recorded and used for subsequent clinical decision-making.

8. The deep learning-driven three-dimensional medical image diagnostic assistance system as described in claim 1, characterized in that: The results display module presents the model's diagnostic predictions to doctors and patients in an intuitive and easy-to-understand manner; it displays the patient's 3D image using a graphical user interface and marks the lesion areas predicted by the model on the image; it also provides a detailed diagnostic report, including the model's prediction probability, disease type, and suggested further examinations and treatment plans; in addition, it uses visualization tools to show the changes in loss and performance indicators during the model training process, so that doctors and researchers can evaluate the model's performance and stability. The goal of the results presentation module is to provide comprehensive and accurate information to help doctors make better clinical decisions.

Images