A deep learning-based general disease prediction model and application thereof

Abstract

The application discloses a general disease prediction model based on deep learning and application thereof, takes lung diseases as an analysis template, applies a MobileNet V2 network architecture, combines a residual structure of ResNet and a gated multimodal fusion unit GMU for image feature extraction, avoids network degradation and overfitting, and ensures high-precision prediction.The model has high universality, can realize prediction of various diseases by replacing a data set, and can also complete a prediction task in the case that part of data is missing under the condition that key data is complete.The model is integrated with a medical work flow through a user-friendly interface, is implanted into a lightweight hardware device RV1126 Linux development board, supports offline use, and improves convenience and operability.The application can accurately predict disease types and severity under the condition that key data is missing, is suitable for primary medical scenes, and can effectively assist doctors in disease diagnosis and treatment decision-making.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and medical diagnostic technology, and in particular to a general disease prediction model based on deep learning and its applications. Background Technology

[0002] With the rapid development of artificial intelligence technology, its application in the medical field is becoming increasingly widespread. However, most existing medical auxiliary diagnostic systems rely on single-modality data or single disease types, lacking versatility and flexibility. Especially in primary healthcare settings, due to limited medical resources, patients often cannot obtain complete medical test data (such as "gold standard" data), which limits the practicality of existing models.

[0003] In existing technologies, deep learning-based medical diagnostic models typically rely on complete medical data and are mostly optimized for specific diseases or data types, making it difficult to make effective predictions when data is missing. For example, existing deep learning models usually require complete CT images, laboratory test results, and other data, while in primary healthcare institutions, patients often can only provide partial data, rendering existing models ineffective.

[0004] Furthermore, existing systems typically require high-performance computing equipment, making them difficult to widely apply in resource-constrained primary healthcare institutions. They also lack versatility, often designed for single diseases (such as pneumonia or gastric cancer), and struggle to capture deep intermodal relationships through simple feature concatenation or weighted fusion. Moreover, they cannot dynamically adapt to data-deficient scenarios or be easily extended to other diseases by simply replacing the dataset. On the hardware side, while toolchains like RKNN support quantized model deployment, they require specific environment configurations and lack optimization solutions for healthcare scenarios.

[0005] Therefore, in order to solve the above-mentioned technical problems, it is an urgent technical problem for those skilled in the art to provide a medical auxiliary diagnostic platform that can make accurate predictions even when key data is missing, and has high versatility and lightweight characteristics. Summary of the Invention

[0006] In view of this, the present invention provides a general disease prediction model based on deep learning and its application, which achieves rapid and accurate diagnosis of lung diseases through the organic combination of front-end user interface, back-end model inference and hardware deployment.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A general disease prediction model based on deep learning includes: a front-end user interface, a back-end model inference, and hardware deployment; wherein, The front-end user interface is developed based on the Electron framework and supports users to upload lung imaging data and input clinical feature data; The backend model inference is developed based on the lightweight deep learning architecture MobileNet V2, which supports the fusion analysis of multimodal data; The hardware is deployed on a hardware platform based on the RK1126 development board, supporting low-power, high-performance real-time inference, and providing a 7-inch, 1080p LCD touchscreen to enhance human-computer interaction capabilities.

[0008] Preferably, the front-end user interface includes the following functional modules: an image upload module, a feature input module, a result display module, and a network configuration module; wherein, The image upload module is used to allow users to select and upload lung CT or X-ray image files through a graphical interface; The feature input module is used by users to input clinical feature data related to lung diseases; The results display module is used to present the diagnostic results to the user in a visual manner; The network configuration module allows users to configure the IP address and port of the backend service through the interface, making it convenient to use in different network environments.

[0009] Preferably, the clinical characteristic data includes age, gender, and routine blood indicators.

[0010] Preferably, the diagnostic results include disease classification, confidence level, and relevant prompts.

[0011] Preferably, the backend model inference implementation includes the following steps: Step 1: Perform data preprocessing The uploaded image data is standardized, including image correction, scaling, and normalization. Step 2: Perform model training For the dataset, data augmentation is performed through random flipping, rotation, and shearing operations to enhance the model's generalization ability; For the model architecture, a pre-trained MobileNet V2 is used as the base model, and the classification of lung diseases is achieved through transfer learning and classifier modification. Step 3: Perform model quantization The trained PyTorch model is converted to RKNN format and quantized to adapt to resource-constrained hardware environments. Step 4: Deploy the hardware Develop a C++-based HTTP inference server to receive image and blood routine data, and perform fusion inference by combining image models and blood routine models.

[0012] Preferably, the hardware deployment includes the following steps: Step a: Perform system compilation Use cross-compilation tools to compile the system and model into executable files that can run on the RK1126 development board; Step b: Implement network configuration The ConnMan daemon can be shut down via a startup script, a fixed IP address and DNS server can be set, and intranet penetration technology is supported. Step c: Perform storage management Allocate independent storage space for projects to avoid conflicts between the system's built-in project library and user projects; Step d: Provide auto-start service The inference service is automatically run when the development board starts by using a self-starting script, ensuring the stability and availability of the system.

[0013] Application of a general disease prediction model based on deep learning in the medical field.

[0014] The present invention achieves the following technical effects compared to the prior art: (1) This invention innovatively utilizes gated multimodal fusion technology, which can dynamically adjust the weight allocation of different modal data during the fusion process based on their real-time quality and correlation. By introducing a gating mechanism, the reliability of each modal data can be intelligently evaluated, assigning higher weights to modal data with complete information and low noise, while automatically reducing the weight of modal data with missing information or high noise. This dynamic weight adjustment strategy not only effectively improves the accuracy and robustness of multimodal data fusion, but more importantly, it enables reliable prediction even when some modal data is missing or incomplete, ensuring the stable operation and effective decision support of the system in complex and non-ideal data environments.

[0015] (2) This invention adopts a highly flexible and scalable general model architecture design. This architecture decouples the core algorithm from the feature data of specific diseases, enabling the system to adapt to the characteristics of datasets for different diseases through standardized data interfaces and modular feature processing units. When a prediction task is required for a new disease, there is no need for large-scale modifications to the overall model architecture and core algorithms; only the corresponding labeled dataset needs to be replaced, and the model can be adapted to the new disease through simple parameter fine-tuning. This design significantly reduces the cost and complexity of model migration and expansion, greatly enhances the system's versatility and multi-disease compatibility, and lays a solid foundation for its widespread application in different medical scenarios.

[0016] (3) This invention places particular emphasis on the feasibility and cost-effectiveness of deployment in practical applications, successfully embedding the core algorithm model into lightweight hardware devices. Specifically, based on the cost-effective RV1126 Linux development board, a hardware platform that enables low-cost offline deployment of the model. The RV1126 development board has efficient edge computing capabilities and optimized energy management, which can meet the needs of the model for real-time data processing and prediction locally, while avoiding dependence on cloud computing resources, effectively protecting data privacy and reducing network transmission costs. Through hardware and software co-optimization, the efficient operation of the model on lightweight hardware is ensured, enabling the technical solution of this invention to be quickly deployed and applied in resource-limited primary healthcare units or mobile healthcare scenarios with low economic investment. Attached Figure Description

[0017] Figure 1 System overall architecture diagram. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention discloses a general disease prediction model based on deep learning. Front-end user interface: The front-end user interface is developed based on the Electron framework and supports cross-platform operation. Users can upload lung image files and input relevant clinical feature data through the graphical interface. The system sends this data to the back-end server for processing and displays the diagnostic results to the user in a visual manner.

[0020] Interface design includes the following steps: File structure: Main process implementation: The main process is responsible for managing the application's lifecycle, creating windows, and handling communication between processes.

[0021] First, import the dependencies: electron -- Uses `app` to manage the application lifecycle, `BrowserWindow` to create windows, and `ipcMain` to handle inter-process communication.

[0022] `path` -- handles file paths to ensure cross-platform compatibility.

[0023] axios -- sends HTTP requests to the backend service.

[0024] Then create a window, load the front-end page index.html, and set it to full-screen by default: Finally, inter-process communication is handled: Listen for the classify-image event (triggered by the rendering process).

[0025] Use axios to send a POST request to serverUrl, carrying image (image data) and features (classification features).

[0026] On success, response data is returned via the classify-result channel; on failure, an error message is returned.

[0027] Rendering process implementation: First, import the dependencies used for communication with the main process: Next, image selection and preview are performed. After the user selects an image, the file is read using FileReader and converted to Base64 format. The element displays a preview image.

[0028] Then, implement form submission, ensure that the user has selected an image, convert it to Base64 format, convert the user-input feature data into floating-point numbers, and default to -1 for uninputted feature data. Finally, send the complete {image, features} object to the main process.

[0029] The classification results are displayed on the user interface, showing the results returned by the server. User interface implementation: The user interface is composed of HTML and CSS, and supports image uploading, feature input, and result display.

[0030] Write code for the page header: Image upload area: The preview image area displays user-uploaded images: Feature data selection area: Select and input some feature data, mainly blood routine data. Backend model section: The backend of this invention is primarily responsible for processing the lung imaging data and clinical feature data uploaded from the frontend, classifying the disease using a deep learning model, and returning diagnostic results. The core functions of the backend include: ① Data preprocessing and augmentation; ② Training and optimization of deep learning models; ③ Quantization and transformation of the model to adapt to hardware deployment; ④ Development and deployment of real-time inference services.

[0031] Project Structure: Data preprocessing: Because raw lung CT images may exhibit perspective distortion, geometric correction is necessary. This patent develops an image correction tool based on OpenCV: To improve the model's generalization ability, the image data is augmented with the following operations: random horizontal flipping, random rotation (±10 degrees), random shearing transformation, random adjustment of brightness and contrast, and normalization.

[0032] Data preprocessing: This invention selects the lightweight MobileNet V2 as the base model and optimizes it by combining it with the residual structure of ResNet: The Adam optimizer was used with learning rates of 0.0001 (feature layer) and 0.001 (classifier layer). Cross-entropy loss was applied, and training was conducted for 50 epochs with a batch size of 4 and a validation set ratio of 20%.

[0033] For model evaluation, the following evaluation metrics are monitored during training: F1 score (weighted average), ROC-AUC score (macro average), confusion matrix, and classification accuracy for each class.

[0034] Hardware components: The system is deployed on the RK1126 development board, supporting low-power, high-performance real-time inference. Hardware deployment includes the following steps: System compilation: Use cross-compilation tools to compile the system and model into executable files that can run on the RK1126 development board.

[0035] Network configuration: This invention shuts down the ConnMan daemon process via a self-starting script and sets a fixed IP address and DNS server.

[0036] Storage Management: Allocate independent storage space for projects to avoid conflicts between the system's built-in project library and user projects.

[0037] Automatic startup service: This invention uses a self-starting script to automatically run the inference service when the development board starts.

[0038]

[0039] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the technical scope of the present invention. Therefore, any minor modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A general disease prediction model based on deep learning, characterized in that, include: Front-end user interface, back-end model inference, and hardware deployment; among which, The front-end user interface is developed based on the Electron framework and supports users to upload lung imaging data and input clinical feature data; The backend model inference is developed based on the lightweight deep learning architecture MobileNet V2, which supports the fusion analysis of multimodal data; The hardware is deployed on a hardware platform based on the RK1126 development board, supporting low-power, high-performance real-time inference, and providing a 7-inch, 1080p LCD touchscreen to enhance human-computer interaction capabilities.

2. The general disease prediction model based on deep learning according to claim 1, characterized in that, The front-end user interface includes the following functional modules: image upload module, feature input module, result display module, and network configuration module; among which, The image upload module is used to allow users to select and upload lung CT or X-ray image files through a graphical interface; The feature input module is used by users to input clinical feature data related to lung diseases; The results display module is used to present the diagnostic results to the user in a visual manner; The network configuration module allows users to configure the IP address and port of the backend service through the interface, making it convenient to use in different network environments.

3. The general disease prediction model based on deep learning according to claim 2, characterized in that, The clinical characteristics data include age, gender, and routine blood count indicators.

4. A general disease prediction model based on deep learning according to claim 2, characterized in that, The diagnostic results include disease classification, confidence level, and relevant information.

5. A general disease prediction model based on deep learning according to claim 1, characterized in that, The specific implementation of the backend model inference includes the following steps: Step 1: Perform data preprocessing The uploaded image data is standardized, including image correction, scaling, and normalization. Step 2: Perform model training For the dataset, data augmentation is performed through random flipping, rotation, and shearing operations to enhance the model's generalization ability; For the model architecture, a pre-trained MobileNet V2 is used as the base model, and the classification of lung diseases is achieved through transfer learning and classifier modification. Step 3: Perform model quantization The trained PyTorch model is converted to RKNN format and quantized to adapt to resource-constrained hardware environments. Step 4: Deploy the hardware Develop a C++-based HTTP inference server to receive image and blood routine data, and perform fusion inference by combining image models and blood routine models.

6. A general disease prediction model based on deep learning according to claim 1, characterized in that, The hardware deployment includes the following steps: Step a: Perform system compilation Use cross-compilation tools to compile the system and model into executable files that can run on the RK1126 development board; Step b: Implement network configuration The ConnMan daemon can be shut down via a startup script, a fixed IP address and DNS server can be set, and intranet penetration technology is supported. Step c: Perform storage management Allocate independent storage space for projects to avoid conflicts between the system's built-in project library and user projects; Step d: Provide auto-start service The inference service is automatically run when the development board starts by using a self-starting script, ensuring the stability and availability of the system.

7. The application of a general disease prediction model based on deep learning as described in claims 1-6 in the medical field.

Images