Automatic sketching system for nasopharyngeal carcinoma tumor

Abstract

The invention discloses a nasopharynx cancer tumor automatic sketching system, which automatically segments a tumor area in image data through a deep learning algorithm, optimizes a segmentation result in combination with a multi-modal image, and provides an accurate three-dimensional visual model and a diagnosis report. Besides, the system supports continuous training and updating of the model, personalized treatment schemes and radiotherapy dose simulation can be generated according to sketching results, and intelligent support is provided for diagnosis and treatment of nasopharyngeal carcinoma. According to the system, the diagnosis and treatment efficiency and accuracy are remarkably improved, and the system has wide clinical application value.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing technology, and specifically to an automatic delineation system for nasopharyngeal carcinoma tumors, which is suitable for the diagnosis, treatment planning and radiotherapy dose design of nasopharyngeal carcinoma, and aims to improve the accuracy of tumor region segmentation and the efficiency of diagnosis and treatment through deep learning technology and multimodal image analysis methods. Background Art

[0002] Nasopharyngeal carcinoma is a common malignant tumor whose diagnosis and treatment rely on accurate medical image analysis. Traditionally, tumor area delineation is mainly completed by manual annotation by radiologists. However, this method is not only time-consuming and labor-intensive, but the annotation results are easily affected by subjective factors, resulting in differences in delineation accuracy. How to achieve efficient and automated delineation of tumor areas has become an urgent problem to be solved.

[0003] At present, automatic segmentation technology based on image processing and machine learning has been initially applied in the field of medical imaging, but existing systems usually rely only on single-modality images and cannot fully utilize the information fusion advantages of multi-modality images (such as MRI, CT, and PET). In addition, the segmentation results of these systems are insufficient in processing small-volume tumors or tumors with irregular boundaries, and the segmentation accuracy and clinical usability need to be improved urgently.

[0004] Existing automatic delineation systems also lack visualization and user interaction functions for segmentation results, and cannot provide intuitive tumor 3D models for physicians to review. At the same time, most systems do not consider the continuous training and optimization capabilities of the model, making it difficult to adapt to dynamic changes in clinical data. Therefore, developing an automatic delineation system for nasopharyngeal carcinoma tumors that combines multimodal image analysis, deep learning algorithms, and 3D visualization technology has important clinical value and application prospects.

[0005] Although some existing automated segmentation systems have made progress in tumor delineation, most systems lack flexibility and cannot update and train models in real time based on the latest imaging data. In addition, many systems lack effective support for quantitative analysis of tumors, staging prediction, and generation of personalized treatment plans. Therefore, establishing an intelligent nasopharyngeal carcinoma tumor automatic delineation system that integrates multimodal imaging, deep learning, and personalized treatment plans can effectively improve the accuracy of delineation and enhance the intelligence of clinical treatment, which has become an urgent problem to be solved in the current field of medical imaging analysis. Summary of the invention

[0006] The present invention provides an automatic delineation system for nasopharyngeal carcinoma tumors, which aims to achieve efficient and accurate delineation of tumor areas in medical images through modular design and deep learning technology, and solve the shortcomings of low efficiency, insufficient precision and lack of flexibility in the prior art. The present invention includes the following modules: data acquisition and preprocessing module, tumor area segmentation module, segmentation result optimization module, three-dimensional visualization and interaction module, result storage and output module, model training and updating module, diagnosis and treatment auxiliary decision module. The following is a detailed introduction to each module:

[0007] Data acquisition and preprocessing module

[0008] The data acquisition and preprocessing module is mainly used to acquire multimodal medical imaging data (including MRI, CT, and PET) and standardize the raw imaging data. First, the system accesses the imaging data source and performs noise filtering to remove artifacts and interference in the image. Secondly, the grayscale range of different modal data is adjusted through grayscale normalization technology so that it can be processed under a unified standard. Finally, the interpolation technology is used to adjust the resolution of the imaging data to adapt to the input requirements of the deep learning algorithm and provide high-quality data input for subsequent processing steps.

[0009] Tumor region segmentation module

[0010] The tumor region segmentation module is the core functional module of the system. It uses a deep learning algorithm based on a convolutional neural network (CNN) to segment the tumor region. By classifying the preprocessed image data pixel by pixel, the module can quickly identify the boundaries of the tumor region. The convolutional neural network architecture adopts an encoder-decoder structure, in which the encoder is responsible for extracting multi-level features in the image, and the decoder is used to generate segmentation results consistent with the input resolution. This module supports large-scale data parallel processing, which can significantly improve the efficiency and accuracy of tumor segmentation.

[0011] Segmentation result optimization module

[0012] The segmentation result optimization module is designed to solve the errors and discontinuities that may occur during the automatic segmentation process. This module combines the characteristics of multimodal images (MRI, CT, and PET) and corrects the initial segmentation results through post-processing algorithms. For example, for areas with complex tumor boundaries or irregular morphology, this module uses an optimization strategy based on image gradients and regional connectivity to ensure that the boundaries of the segmentation results are accurate and the regions are complete. In addition, this module also uses the correction data annotated by doctors for feedback iteration, thereby improving the overall reliability of the system.

[0013] 3D visualization and interaction module

[0014] The 3D visualization and interaction module provides an intuitive way to display the tumor area. Through the 3D reconstruction algorithm, the module generates a 3D tumor model based on the segmentation results and supports real-time user interaction, including model rotation, scaling, slice browsing and annotation functions. Doctors can use this module to intuitively understand the morphological characteristics of the tumor in different planes and dimensions, thereby verifying and adjusting the automatic segmentation results and further improving the accuracy of diagnosis.

[0015] Result storage and output module

[0016] The result storage and output module is used to save the segmentation results and diagnostic data generated by the system. The segmentation results are stored in DICOM format and are compatible with the hospital information system (HIS) and radiotherapy planning system (TPS). In addition, the module generates standardized reports containing tumor volume, location, and segmentation accuracy assessment to provide comprehensive support for subsequent treatment. Doctors and researchers can intuitively understand the segmentation effect and lesion characteristics through the report, which is convenient for archiving and sharing.

[0017] Model training and updating module

[0018] The model training and update module optimizes system performance through a continuous learning mechanism. This module uses transfer learning and data enhancement technology to fine-tune and expand the original model using newly added imaging data and correction feedback from doctors. During the transfer learning process, the model can be adapted according to different patient characteristics, imaging equipment, and data distribution, thereby ensuring its stable performance in diverse clinical scenarios.

[0019] Diagnosis and treatment decision support module

[0020] The diagnosis and treatment decision-making module generates personalized treatment plans and decision support content based on the segmentation results. By simulating the distribution of radiotherapy doses, the module provides recommendations for the best treatment plan for patients, including tumor target coverage and protection strategies for surrounding normal tissues. In addition, combined with the historical lesion change trend and segmentation result analysis, the module can predict the direction of tumor development, provide data support for doctors' treatment planning, and improve treatment effects.

[0021] The automatic delineation system for nasopharyngeal carcinoma tumors completes the entire process from image data acquisition and preprocessing to automatic tumor segmentation, generation of personalized treatment plans, and radiotherapy dose simulation through the collaborative work of multiple modules.

[0022] Image data preprocessing module: First, the original image data is processed to ensure that the subsequent processing can achieve better results. After processing by this module, the image quality is improved, the noise is removed, the tumor area is more obvious, and clear input data is provided for the tumor segmentation module.

[0023] Tumor segmentation module: Deep learning technology is then used to automatically identify the tumor area and outline the precise boundaries of the tumor, providing basic data for the generation of treatment plans.

[0024] Multimodal image fusion module: The image data of different modalities are then effectively fused to further optimize the segmentation results of the tumor area and improve the overall segmentation accuracy of the system.

[0025] Model training and updating module: It ensures that the segmentation model can be continuously optimized and maintain high accuracy by providing real-time feedback of new case data.

[0026] Treatment plan generation module: Based on tumor segmentation results and individual patient characteristics, it intelligently generates personalized treatment plans to provide doctors with scientific treatment decision support.

[0027] Radiotherapy dose simulation module: By simulating the distribution of radiotherapy dose, it helps doctors achieve accurate irradiation during radiotherapy, ensure the effectiveness of treatment, and reduce damage to healthy tissues.

[0028] Beneficial Effects

[0029] The automatic delineation system for nasopharyngeal carcinoma tumors provided by the present invention has the following beneficial effects:

[0030] Significantly improve tumor delineation efficiency

[0031] Through the introduction of deep learning algorithms, the system realizes the automatic segmentation of tumor areas in imaging data, avoiding the time-consuming and subjective problems of traditional manual outlining, and greatly improving clinical work efficiency.

[0032] Accurate multimodal data fusion

[0033] The system makes full use of multimodal imaging data such as MRI, CT and PET. Through the collaborative work of the data acquisition and preprocessing module and the segmentation result optimization module, it effectively solves the problem that a single modality image is difficult to accurately describe the tumor morphology and boundaries, thereby improving the accuracy and robustness of tumor segmentation.

[0034] Enhance doctors’ interactive capabilities

[0035] Through the 3D visualization and interaction module, doctors can intuitively view the 3D morphology of the tumor area, quickly verify the segmentation results and make necessary adjustments, providing flexible support for clinical diagnosis and treatment plan formulation.

[0036] Intelligent diagnosis and treatment planning

[0037] The system uses the diagnosis and treatment decision-making support module to generate personalized treatment plans and radiotherapy dose distribution simulations based on the results of automatic outlining, which helps to formulate accurate radiotherapy plans, reduce damage to normal tissues, and optimize treatment effects.

[0038] Support dynamic optimization and learning

[0039] The model training and update module supports the continuous optimization of system performance by introducing new imaging data and doctor feedback, ensuring that it can adapt to different patient data and dynamically changing clinical needs.

[0040] Improve clinical promotion and use value

[0041] Through the result storage and output module, the segmentation results are saved in DICOM format and a standardized report is generated, which is seamlessly connected to the existing medical system, making it convenient for doctors to use and archive. At the same time, the system's automation and standardized processing flow lowers the threshold for use, making it easier to promote and apply it in medical institutions.

[0042] Main innovations:

[0043] Multimodal imaging tumor segmentation based on deep learning

[0044] This paper uses convolutional neural networks (CNN) to jointly analyze multimodal medical images (MRI, CT and PET), and for the first time applies multimodal image fusion to the automatic segmentation of nasopharyngeal carcinoma tumor areas. Through convolution feature extraction, pixel-by-pixel classification and feature level fusion, this method significantly improves the ability to recognize complex tumor boundaries and irregular morphology. Compared with traditional single-modal image segmentation methods, the innovative design of multimodal fusion makes up for the problem of insufficient single-modal image information and greatly improves the accuracy and robustness of tumor segmentation.

[0045] Dynamic post-processing mechanism for segmentation result optimization

[0046] The system innovatively designed a segmentation result optimization module, which solves the discontinuity and boundary fuzziness problems that may occur in the tumor region in traditional automatic segmentation algorithms by dynamically post-processing the initial segmentation results. The optimization mechanism uses multimodal imaging features to correct tumor boundaries and combines regional connectivity analysis algorithms to enhance the recognition of small-volume tumors, thereby significantly improving the integrity and accuracy of tumor region segmentation.

[0047] Integration of 3D visualization and interactive functions

[0048] The system is designed with a 3D visualization and interaction module, which generates an intuitive 3D model from the tumor segmentation results and provides real-time interactive functions, including model rotation, zooming, and slice browsing. This innovative function enhances doctors' ability to understand and verify segmentation results, provides an operational reference interface for clinical diagnosis and treatment plan formulation, and makes up for the lack of intuitive visualization functions in traditional systems.

[0049] Diagnosis and treatment decision support and personalized treatment planning

[0050] The present invention uses a diagnosis and treatment decision-making auxiliary module to generate personalized treatment plans and radiotherapy dose distribution simulations based on automatic segmentation results, directly linking tumor segmentation results with treatment planning for the first time. The module uses the radiotherapy dose distribution model to optimize tumor target coverage, reduce damage to normal tissues, and predict tumor development trends, providing doctors with scientific and accurate treatment recommendations, significantly improving the level of intelligent diagnosis and treatment.

[0051] Support dynamic learning and continuous model optimization

[0052] The system has designed a model training and update module, and by introducing transfer learning and data enhancement technology, the system can dynamically adapt to new imaging data and doctor correction feedback. This innovation solves the defects of traditional systems in which model performance is fixed and it is difficult to cope with data diversity, ensuring the long-term applicability and high performance of the system in diverse clinical scenarios.

[0053] Full process standardization and efficient integrated design

[0054] The present invention forms a complete standardized processing flow from data acquisition, preprocessing to segmentation result optimization, three-dimensional visualization, diagnosis and treatment assistance, and result storage. It innovatively combines multimodal image analysis, deep learning algorithms, and clinical workflows, reduces operational complexity, significantly improves diagnosis and treatment efficiency, and lays a solid foundation for the promotion and application of the system in medical institutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is the flow chart of data acquisition and preprocessing module;

[0056] Figure 2 It is the flow chart of the tumor region segmentation module;

[0057] Figure 3 It is the flow chart of the segmentation result optimization module;

[0058] Figure 4 It is a flow chart of the 3D visualization and interaction module;

[0059] Figure 5 It is the flow chart of result storage and output module;

[0060] Figure 6 It is the flow chart of model training and updating module;

[0061] Figure 7 It is the flow chart of diagnosis and treatment auxiliary decision-making module. DETAILED DESCRIPTION

[0062] The embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

[0063] Data acquisition and preprocessing module

[0064] The data acquisition and preprocessing module is the first step of the present invention, which is mainly responsible for collecting medical imaging data of nasopharyngeal carcinoma patients from a variety of imaging devices (such as MRI, CT, PET, etc.) and performing unified preprocessing on the data.

[0065] Data acquisition: After the patient is scanned by CT, MRI and PET scanners, the image data is formatted and uploaded to the system through the DICOM protocol. Each type of image data contains different tissue information. For example, MRI images provide detailed structure of soft tissue, CT images can clearly show bone tissue, and PET images provide information on the metabolic activity of the tumor. By integrating multimodal data, the system can identify the tumor area more comprehensively and accurately.

[0066] Data preprocessing: The system standardizes the collected image data. First, the grayscale normalization method is used to adjust the grayscale value range of different images (for example, the grayscale values ​​of all images are unified to between 0 and 1) to ensure data consistency. Then, the Gaussian filtering algorithm is used to remove noise from the image, smooth the image content, and prevent noise from interfering with the subsequent analysis process. Finally, the system will unify the resolution of the image through the interpolation method, and adjust the images of different resolutions to 256×256 pixels to adapt to the deep learning model.

[0067] Tumor area automatic segmentation module

[0068] After the data acquisition and preprocessing module completes data processing, the processing results enter the tumor area automatic segmentation module, and the tumor area is accurately segmented using a deep learning algorithm.

[0069] Deep learning model: This module uses convolutional neural networks (CNN), especially the U-Net model, to automatically segment tumor areas. The U-Net model first extracts high-dimensional features of the image (such as the edge and texture of the tumor) through its encoder and decoder structure, and then generates a segmentation map of the same size as the input image by pixel-by-pixel classification. Each pixel in the map is labeled as a "tumor" or "non-tumor" area.

[0070] Tumor identification and boundary extraction: In the initial segmentation map generated by the model, the boundary of the tumor area may not be completely clear or there may be some errors. Therefore, the system uses gradient information to further optimize the tumor boundary, ensure the accuracy of the tumor area boundary in the segmentation map, and remove smaller, non-tumor areas.

[0071] This module uses cross entropy loss weighting during model training to improve segmentation accuracy and solve the problem of class imbalance. The specific implementation method is as follows:

[0072]

[0073] in:

[0074] L: weighted cross entropy loss;

[0075] y i : True label, value is 0 or 1;

[0076] p i : predicted probability;

[0077] ω1, ω2: Category weighting coefficients, with a value range of [0,1], used to adjust the weights of tumor and non-tumor categories.

[0078] Segmentation result optimization module

[0079] The segmentation result optimization module further optimizes the preliminary results produced by automatic segmentation to improve the segmentation accuracy and usability of the results.

[0080] Multimodal image fusion: This module first fuses the MRI, CT, and PET images. By jointly analyzing these images, the system can combine the advantages of each image to optimize the segmentation results. For example, the soft tissue resolution provided by MRI images can supplement the details lost in CT images, while PET images can provide accurate information about the metabolic areas of the tumor. By combining these modalities, the system can more accurately identify the boundaries and internal features of the tumor.

[0081] Regional connectivity analysis and repair: In order to avoid mis-segmentation and blurred boundaries, the optimization module uses a regional connectivity analysis algorithm. This algorithm checks whether each region in the segmentation result is coherent. If it is detected that the tumor region is mis-segmented into multiple disconnected parts, the system will automatically merge them. At the same time, if there are missed small tumors or blurred areas, the system will complete them through the expansion algorithm.

[0082] 3D visualization and interaction module

[0083] The optimized segmentation results will be transmitted to the 3D visualization and interaction module, which displays the spatial distribution of the tumor through 3D reconstruction and provides a convenient interactive method for doctors to analyze and confirm.

[0084] 3D model generation: The system uses the Marching Cubes algorithm to convert the optimized tumor area segmentation results into a 3D surface model. This 3D model can truly reflect the spatial position, shape and size of the tumor, helping doctors to understand the specific situation of the tumor more intuitively.

[0085] Real-time interactive function: Doctors can rotate, zoom, and slice the 3D model in real time through the system interface. By observing from different perspectives, doctors can more clearly understand the shape and location of the tumor and further verify the accuracy of the automatic segmentation results.

[0086] Marking and annotation function: Doctors can also add annotations to the 3D model, such as marking special areas of the tumor, measuring tumor volume, or marking suspected metastatic areas. This function enables doctors to make personalized diagnoses more flexibly and record important diagnostic information to facilitate subsequent treatment decisions.

[0087] Result storage and output module

[0088] The optimized and verified segmentation results and 3D visualization models will be transmitted to the result storage and output module.

[0089] Storage format: The segmentation results will be saved in the standard DICOM format to ensure compatibility with the hospital's existing information systems (such as HIS and PACS systems). In this way, doctors can easily access, manage and share patients' imaging data.

[0090] Report generation: The system automatically generates standardized reports that contain detailed data on tumor segmentation results, such as tumor volume, maximum cross-sectional area, segmentation accuracy (such as Dice coefficient), 3D model screenshots, and radiotherapy dose distribution. These reports can be exported in PDF format for doctors' reference and clinical decision-making.

[0091] Model training and updating module

[0092] This module continuously updates and optimizes the system's segmentation model to respond to new patient data and changing clinical needs.

[0093] Incremental learning: Each time the system is used, doctors may modify the automatic segmentation results, especially when the tumor morphology is irregular or the image quality is poor. The system uses these modifications together with the original data as new training samples, and fine-tunes the deep learning model through the incremental learning mechanism to improve its performance on similar data.

[0094] Performance evaluation and update: After each model update, the system evaluates the segmentation accuracy of the model through cross-validation and optimizes it when necessary to ensure that the system can provide more accurate tumor segmentation results on new datasets.

[0095] Diagnosis and treatment decision support module

[0096] Based on tumor segmentation and three-dimensional reconstruction, the system provides doctors with personalized treatment plans through the diagnosis and treatment auxiliary decision-making module.

[0097] Personalized treatment plan generation: Based on the automatically segmented tumor area and the patient's specific situation, the system combines the existing medical database and radiotherapy plan to generate a personalized treatment plan. The system considers the patient's tumor size, location and clinical characteristics to recommend a suitable radiotherapy plan.

[0098] Radiotherapy dose distribution simulation: The system automatically calculates the radiotherapy dose distribution based on the morphology and location of the tumor and generates a simulation diagram. Radiotherapy dose simulation helps doctors estimate the treatment effect before treatment and optimize the dose distribution to ensure that the tumor area is fully irradiated while avoiding damage to surrounding normal tissues.

Claims

1. An automatic delineation system for nasopharyngeal carcinoma tumors, characterized in that: include: Data acquisition and preprocessing module, used to collect and process nasopharyngeal medical imaging data and perform standardized processing; Tumor region segmentation module, which automatically segments tumor regions in image data through deep learning algorithms; The segmentation result optimization module combines multimodal images to correct and optimize the segmentation results and improve the segmentation accuracy; 3D visualization and interaction module, used to generate 3D visualization models of tumors and support user interactive adjustment and annotation; Result storage and output module, used to save the final delineation results and generate standardized diagnostic reports; Model training and updating module, which updates deep learning models and improves model performance by continuously inputting new data and providing feedback results; The diagnosis and treatment decision-making support module generates personalized treatment plans based on the delineation results, and provides radiotherapy dose distribution simulation and lesion trend prediction.

2. The system according to claim 1, characterized in that Including, the data acquisition and preprocessing module is used to collect nasopharyngeal medical imaging data including MRI, CT and PET, and perform standardized processing through noise filtering, grayscale normalization and resolution adjustment.

3. The system according to claim 1, characterized in that Including, the tumor region segmentation module automatically segments the tumor region in the image data based on a convolutional neural network (CNN) to achieve high-precision tumor recognition, and the segmentation loss function adopts a cross-entropy loss weighted form, and the calculation formula is: Among them, L: weighted cross entropy loss, y i is the true label of the i-th pixel (the actual classification or belonging category of the sample), p i is the prediction probability (the confidence output by the model about whether the sample belongs to a certain category, which is the prediction result of the model), ω1 and ω2 are category weighting coefficients (used to adjust the weights of different categories in model training, in order to solve the problem of category imbalance).

4. The system according to claim 1, characterized in that Including, the segmentation result optimization module combines multimodal image data and rule-based post-processing algorithms to correct and optimize discontinuous areas in the segmentation results, thereby improving the accuracy and consistency of the segmentation results.

5. The system according to claim 1, characterized in that Including, the three-dimensional visualization and interaction module uses a three-dimensional reconstruction algorithm to generate a three-dimensional model of the tumor area, and supports users to perform interactive operations such as model rotation, scaling and slice browsing.

6. The system according to claim 1, characterized in that Including, the result storage and output module supports saving the segmentation results in DICOM format and generating a standardized report containing tumor volume, location and diagnostic information.

7. The system according to claim 1, characterized in that Including, the model training and updating module optimizes and updates the deep learning model by continuously introducing new medical imaging data and user feedback, and using transfer learning and data enhancement technology.

8. The system according to claim 1, characterized in that Including, the diagnosis and treatment auxiliary decision module generates a personalized treatment plan according to the tumor delineation results, including radiotherapy dose distribution simulation and tumor lesion growth trend prediction, to assist medical diagnosis and treatment planning.

Images