Artificial intelligence assisted early tumor screening image analysis system

Abstract

The present application relates to the field of medical image processing, in particular to an artificial intelligence assisted early tumor screening image analysis system, comprising: a data preprocessing module for normalizing chest CT, breast mammography X-ray and cervical TCT examination medical images and unifying them into DICOM format; an image segmentation module for segmenting suspicious disease sites based on differential geometry feature analysis, including a manifold feature extraction unit, a curvature analysis enhancement unit and a geodesic flow attention fusion unit; a disease classification module for judging disease types using a ResNet network structure; and a risk assessment module for assessing the risk of suspicious lesions through X-GBoost and residual model; the present application combines differential geometry theory and deep learning technology, effectively solves the technical problems in early detection of small tumors through manifold representation, curvature analysis and geodesic flow attention mechanism, significantly improves the small target detection rate and reduces the false positive rate.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, specifically to an artificial intelligence-assisted early tumor screening image analysis system, which is applied to the fields of medical image-assisted diagnosis and early tumor detection. Background Technology

[0002] Early cancer screening is crucial for improving patient survival rates and quality of life. Currently, commonly used clinical cancer screening methods include chest CT scans, mammograms, and cervical transcranial Doppler (TCT) scans. However, traditional medical image diagnosis relies heavily on physicians' professional experience, facing problems such as low efficiency, high subjectivity, and high rates of misdiagnosis and missed diagnosis due to manual image reading. This is especially true for early, small lesions, which are more easily overlooked due to their indistinct features, blurred boundaries, and low contrast with normal tissue.

[0003] With the development of artificial intelligence (AI) technology, deep learning algorithms have shown broad application prospects in medical image analysis. Current AI-assisted diagnostic systems mainly employ convolutional neural networks for image feature extraction and classification. However, these methods still face the following challenges when processing early-stage tumor images: firstly, limited detection capability for small lesions (typically less than 1% of the image area); secondly, insufficient accuracy in segmenting lesions with blurred boundaries; and thirdly, poor sensitivity to low-contrast lesions. These problems severely restrict the effective application of AI technology in early tumor screening.

[0004] Therefore, there is an urgent need to develop an intelligent image analysis system that can accurately identify and analyze small lesions in early-stage tumors. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-assisted image analysis system for early tumor screening. By introducing an innovative method that combines differential geometry theory with deep learning technology, it solves the technical bottleneck of existing technologies in the detection of early small lesions and improves the accuracy and effectiveness of early tumor screening.

[0006] This invention proposes an artificial intelligence-assisted early tumor screening image analysis system, comprising:

[0007] The data preprocessing module is used to normalize medical images and unify them into DICOM format;

[0008] An image segmentation module, connected to the data preprocessing module, is used to receive a normalized medical image sent by the data preprocessing module and segment suspicious disease areas in the normalized medical image based on differential geometric feature analysis. The image segmentation module includes a manifold feature extraction unit, a curvature analysis enhancement unit, and a geodesic flow attention fusion unit.

[0009] The disease classification module, connected to the image segmentation module, is used to receive the segmentation results from the image segmentation module, determine the disease type from the segmentation results, and generate suspicious lesion detection results.

[0010] A risk assessment module, connected to the image segmentation module, is used to receive the segmentation results from the image segmentation module.

[0011] Preferably, the manifold feature extraction unit is used for:

[0012] The normalized medical image is mapped to a two-dimensional manifold representation;

[0013] Calculate the metric tensor and curvature information on the two-dimensional manifold representation;

[0014] Construct a multi-scale feature pyramid to extract geometric features at different scales.

[0015] Generate a feature map containing local geometric structure information.

[0016] Preferably, the curvature analysis enhancement unit is used for:

[0017] Potential abnormal regions are identified based on the Gaussian curvature and mean curvature in the feature map.

[0018] Construct curvature feature maps to highlight potential lesion areas;

[0019] The enhancement threshold is dynamically determined based on the distribution of the curvature feature map;

[0020] Adaptive geometric enhancement is applied to the identified potential lesion areas;

[0021] Output the enhanced feature map.

[0022] Preferably, the geodesic flow attention fusion unit is used for:

[0023] Define geodesic flows on characteristic manifolds to characterize the evolution of features;

[0024] Identifying key characteristic regions based on geodesic flow evolution patterns;

[0025] A three-layer feature fusion network is constructed, including a feature extraction layer, a feature selection layer, and a fusion output layer;

[0026] Integrating multi-level features through geodesic flow-guided attention mechanisms;

[0027] Generate the final segmentation result.

[0028] Preferably, the data preprocessing module is specifically used for:

[0029] Medical images from chest CT scans, mammograms, and cervical TCT examinations will be uniformly converted to DICOM format.

[0030] Different types of medical images undergo separate data standardization processes, including grayscale normalization, contrast adjustment, and size scaling.

[0031] The medical images are scaled to make them suitable for subsequent model analysis.

[0032] Preferably, the three-layer feature fusion network includes:

[0033] The feature extraction layer includes residual connection units, convolutional feature units, and geometric feature units;

[0034] The feature filtering layer includes a geodesic flow-based attention control unit and a dual-path processing unit, wherein the dual-path processing unit includes a shallow path for preserving detailed information and a deep path for extracting semantic information;

[0035] The fusion output layer includes a feature integration unit, a segmentation prediction unit, and a geometric refinement unit.

[0036] Preferably, the disease classification module adopts a ResNet network structure, which extracts image features and outputs image classification results through a network composed of convolutional layers and pooling layers.

[0037] Preferably, the risk assessment module includes:

[0038] The annotation unit is used to mark the regions of suspicious lesions in the segmentation results and generate the gold standard;

[0039] The risk model construction unit is used to construct X-GBoost and residual models to calculate the risk probability of the suspicious lesions.

[0040] Evaluation unit, used to evaluate the performance of risk prediction models using ROC curves.

[0041] Preferably, the system further includes:

[0042] An image acquisition unit is used to acquire medical images of patients through medical imaging equipment;

[0043] An edge computing unit, connected to the image acquisition unit, is used to receive medical images acquired by the image acquisition unit and perform local image processing and analysis.

[0044] The cloud storage / cloud service unit is connected to the edge computing unit and is used to provide cloud-based data storage, artificial intelligence computing, and remote diagnosis and treatment services.

[0045] Preferably, the geodesic flow attention fusion unit is optimized using a focus loss function during training, with a loss weight set to 0.25. It also employs a multi-stage training strategy to first train each sub-network and then perform end-to-end fine-tuning.

[0046] The beneficial effects of this invention include:

[0047] 1. By extracting manifold features based on differential geometry theory, geometric structural information in medical images can be captured more accurately, especially the morphological features of small lesions, thereby increasing the detection rate of early tumors by 35% to 50%.

[0048] 2. Curvature analysis enhancement technology can effectively identify and enhance low-contrast potential abnormal areas in medical images, significantly improving sensitivity to small lesions;

[0049] 3. By using a geodesic flow-driven attention fusion network, the problem of attention dispersion in small target detection by traditional attention mechanisms is solved, and accurate localization of tiny lesions is achieved, with the boundary localization error reduced to an average of 2-3 pixels;

[0050] 4. The system integrates functional modules such as data preprocessing, image segmentation, disease classification, and risk assessment, providing an end-to-end solution from medical image input to risk assessment output, reducing the false positive rate by more than 40% and reducing unnecessary further examinations;

[0051] 5. The distributed computing architecture based on Docker technology supports large-scale data and model training, improving the system's processing efficiency and scalability, with the processing time for a single CT sequence controlled within 20 seconds. Attached Figure Description

[0052] Figure 1 This is a block diagram of the overall structure of the AI-assisted early tumor screening image analysis system of the present invention;

[0053] Figure 2 This is a flowchart of the data preprocessing module of the present invention;

[0054] Figure 3 This is a structural block diagram of the image segmentation module of the present invention;

[0055] Figure 4 This is a flowchart of the curvature analysis enhancement unit of the present invention;

[0056] Figure 5 This is a flowchart of the risk assessment module of the present invention. Detailed Implementation

[0057] Please refer to Figures 1-5 The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0058] like Figure 1 As shown, this invention provides an artificial intelligence-assisted early tumor screening image analysis system, including a data preprocessing module 1, an image segmentation module 2, a disease classification module 3, and a risk assessment module 4. These modules work together to form a complete processing flow from medical image input to risk assessment output.

[0059] Data preprocessing module 1 is used to normalize different types of medical images and convert them uniformly into DICOM format. For example... Figure 2 As shown, the processing flow of data preprocessing module 1 includes two steps: image format unification and image scaling and standardization. In this embodiment, data preprocessing module 1 can process various medical images such as chest CT scans, mammograms, and cervical TCT scans. For example, mammogram images provided by the breast department of a hospital may be acquired using equipment from different manufacturers, resulting in different formats. Data preprocessing module 1 first converts these images into the standard DICOM format for easier subsequent processing.

[0060] For mammogram X-ray images, the preprocessing module performs gridding and region-of-interest (ROI) annotation on the X-ray images, dividing the breast tissue in each X-ray image into n×n grids, with each grid corresponding to the coordinates of the upper left corner of the image. Preferably, n ranges from 8 to 16, because in clinical practice, grids that are too small (e.g., 4×4) cannot capture sufficient local features, while grids that are too large (e.g., 32×32) increase the computational burden and may introduce excessive noise. For example, with n=12, a 2048×2048 pixel mammogram image will be divided into 144 grids, each approximately 170×170 pixels, suitable for capturing early lesion features such as microcalcifications.

[0061] For chest CT images, the preprocessing module performs data transformation on the original image through random horizontal flipping, random vertical flipping, and random Gaussian transform. Then, it standardizes the image through random grayscale stretching, random contrast adjustment, and random brightness adjustment. The random operation parameters are set as follows: the standard deviation of the Gaussian transform is between 0.1 and 1.0, the grayscale stretching factor is between 0.8 and 1.2, the contrast adjustment factor is between 0.85 and 1.15, and the brightness adjustment value is between -10 and 10. These parameter ranges are based on statistical analysis of a large number of clinical CT images and effectively balance data enhancement effects with image realism.

[0062] Image segmentation module 2 is connected to data preprocessing module 1 and is used to receive the preprocessed normalized medical image and segment suspicious disease areas in the image based on differential geometric feature analysis. For example... Figure 3As shown, the image segmentation module 2 includes a manifold feature extraction unit 21, a curvature analysis enhancement unit 22, and a geodesic flow attention fusion unit 23.

[0063] The manifold feature extraction unit 21 is one of the key innovations of this invention. Its working principle is as follows: this unit treats the medical image as a two-dimensional manifold embedded in a high-dimensional space, and captures the structural information of the image by calculating the geometric properties on the manifold. Specifically, the manifold feature extraction unit 21 first maps the normalized medical image into a two-dimensional manifold representation, then calculates the metric tensor and curvature information on the manifold representation, constructs a multi-scale feature pyramid, and finally generates a feature map containing local geometric structure information.

[0064] In this example, the medical image I is considered as a two-dimensional manifold M embedded in a high-dimensional space, and a correspondence is established through the mapping function φ: I→M. For each point p on the manifold M, a local coordinate system and a Riemannian metric tensor g can be defined. Based on this metric tensor, various curvature information at point p can be calculated, including the Gaussian curvature K and the mean curvature H:

[0065] ,

[0066] in, Let p be the Gaussian curvature at point p, representing the intrinsic geometric properties of that point; and is the principal curvature at point p, representing the degree of curvature at that point in two mutually perpendicular directions; It is an image The gradient vector at point p has a magnitude of This represents the rate of change of the image at that point; the squared term in the denominator ensures the scale invariance of the curvature calculation. In tumor images, Gaussian curvature can effectively distinguish the boundary regions between lesions and normal tissue, because these regions typically exhibit "saddle points" (with negative Gaussian curvature).

[0067] ,

[0068] in, Let be the mean curvature at point p, representing the intrinsic geometric properties of that point; the numerator is the arithmetic mean of the two principal curvatures; the 3 / 2 power term in the denominator ensures the dimensional consistency of the mean curvature. In tumor image analysis, the mean curvature can highlight morphological changes in tissues, especially for early-stage round or elliptical tumors, whose boundaries are usually represented by closed curves with a high mean curvature.

[0069] Taking lung nodule detection in chest CT images as an example, when analyzing a set of 512×512×64 chest CT sequences, the manifold feature extraction unit 21 first maps each slice to a manifold representation, and then calculates the Gaussian curvature and mean curvature of each voxel point. Compared with traditional methods, this feature extraction method based on differential geometry can more accurately identify early lung nodules with a diameter of less than 5 mm because it captures subtle geometric changes in the nodule boundary that may not be obvious in conventional grayscale or texture features.

[0070] To capture geometric features at different scales, the manifold feature extraction unit 21 constructs a feature pyramid structure containing nine scale layers. Each scale layer processes the image through filters of different sizes to extract geometric features at the corresponding scale. Smaller scale layers (scales 1-3) capture pixel-level local geometry, medium-scale layers (scales 4-6) characterize the morphological features of tissue regions, and larger scale layers (scales 7-9) describe organ-level anatomical relationships. Preferably, the filter size for small-scale layers is 3×3 pixels, for medium-scale layers it is 5×5 pixels, and for large-scale layers it is 7×7 or 9×9 pixels. In practical applications, for example, for a mammogram X-ray image, small-scale layers can capture the geometric features of tiny calcifications (typically 0.1–0.5 mm), medium-scale layers can characterize glandular structural changes (1–5 mm), and large-scale layers can describe the overall breast tissue architecture (>10 mm).

[0071] Curvature analysis enhancement unit 22 is another innovation of this invention, and its processing flow is as follows: Figure 4 As shown, this unit utilizes Gaussian curvature and mean curvature to identify potential abnormal regions in the image, constructing a curvature feature map to highlight potential lesion regions. Specifically, the curvature analysis enhancement unit 22 dynamically determines the enhancement threshold based on the curvature distribution in the feature map, performs adaptive geometric enhancement on the identified potential lesion regions, and outputs the enhanced feature map.

[0072] In this embodiment, the curvature analysis enhancement unit 22 first calculates the Gaussian curvature K and the average curvature H at each point on the feature map, and then determines the dynamic threshold τ based on the statistical characteristics of the curvature distribution:

[0073] ,

[0074] in, The dynamic threshold for Gaussian curvature; It is the mean of the Gaussian curvature of all points in the image; It is the standard deviation of Gaussian curvature; It is an adjustable parameter. Preferably, The value ranges from 1.5 to 2.5. For different types of medical images, it can be adjusted according to the image characteristics. Value. For example, for high-contrast CT images, A smaller value (e.g., 1.5) can be used; for ultrasound images with low contrast, A larger value (e.g., 2.5) can be chosen. In practical applications, such as the detection of early-stage liver cancer in liver CT images, the setting of the Gaussian curvature threshold directly affects the detection rate and false positive rate. Setting it to 2.0 achieves a better balance between the two.

[0075] ,

[0076] in, The dynamic threshold for the average curvature; It is the mean of the average curvature of all points in the image; It is the standard deviation of the mean curvature; It is an adjustable parameter. Preferably, The value ranges from 1.2 to 2.0, and can also be adjusted according to different medical image types. For abnormal cell detection in cervical TCT images, the average curvature threshold is usually set relatively low ( =1.2) to capture subtle changes in cell morphology; while for mass detection in mammograms, a higher threshold can be set ( (≈2.0), reducing interference with normal glandular tissue.

[0077] For regions exceeding the threshold, curvature analysis enhancement unit 22 applies a geometric enhancement function E to enhance its feature representation:

[0078] ,

[0079] in, The enhanced eigenvalue at point p; It is the value of the original feature map at point p; and These represent the absolute values ​​of the Gaussian curvature and the mean curvature deviating from the mean, respectively. It is the strength enhancement coefficient. Preferably, The value ranges from 2.0 to 4.0, depending on the image contrast and noise level. For early lung nodule detection, this enhancement function can significantly improve the contrast of small nodules with a diameter of less than 5 mm. For example, for an early lung nodule with an original contrast of only 1.1:1, applying this enhancement function will significantly improve the contrast of the nodule. =3.0), its contrast ratio can be increased to 1.6:1, which greatly improves the detection probability.

[0080] The geodesic flow attention fusion unit 23 is the third innovation of this invention. Its structure is as follows: this unit defines geodesic flows on the feature manifold, characterizes the feature evolution process, and identifies key feature regions based on the geodesic flow evolution law. Specifically, the geodesic flow attention fusion unit 23 constructs a three-layer feature fusion network, integrating multi-level features through a geodesic flow-guided attention mechanism to generate the final segmentation result.

[0081] In this embodiment, the geodesic flow attention fusion unit 23 first defines a geodesic distance metric d on the feature manifold to characterize the semantic correlation between different features:

[0082] ,

[0083] in, Represent two points on the characteristic manifold and Geodetic distance between them; This means taking all connections. and curve The minimum value in; It is a curve with parameter t. ; It is the tangent vector of the curve at point t; It is a point on the curve The Riemannian metric tensor at that location; Indicating in measurement Below, vector The length of the geodesic distance. This geodesic distance metric is of great significance in tumor image analysis, as it reflects semantic distance in feature space. For example, in CT images of lung nodules, feature points within the nodule's interior typically have small geodesic distances, while feature points between the nodule and surrounding normal lung tissue have larger geodesic distances. This helps to more accurately define nodule boundaries.

[0084] Based on geodesic distance, geodesic flow attention fusion unit 23 constructs a geodesic flow field to describe the evolution of features:

[0085] ,

[0086] in, Representation of features The partial derivative with respect to time t describes the rate of change of the feature over time; It is the initial feature; It is a gradient operator based on the Riemannian metric g; It is a feature With initial features The partial differential equation describes the process of a feature moving along a geodesic line on a manifold towards the initial feature, similar to the heat conduction equation, but taking into account the geometry of the manifold. In the analysis of breast tumor MRI images, by solving this equation, the initial feature point can be set at the center of the suspected lesion, and then the evolution of the feature can be observed, which can more accurately determine the extent and boundary of the tumor, especially for non-invasive ductal carcinoma in situ (DCIS) with indistinct boundaries.

[0087] The three-layer feature fusion network of the geodesic flow attention fusion unit 23 includes a feature extraction layer, a feature selection layer, and a fusion output layer. The feature extraction layer includes residual connection units, convolutional feature units, and geometric feature units to extract different types of features. The feature selection layer includes a geodesic flow-based attention control unit and a dual-path processing unit, where the dual-path processing unit includes a shallow path for preserving detailed information and a deep path for extracting semantic information. The fusion output layer includes a feature integration unit, a segmentation prediction unit, and a geometric refinement unit to integrate multi-level features and generate the final segmentation result.

[0088] During training, the geodesic flow attention fusion unit 23 is optimized using a focal loss function with a loss weight of 0.25. A multi-stage training strategy is employed to first train each sub-network and then perform end-to-end fine-tuning.

[0089] ,

[0090] in, The value of the focus loss function; It is the model's predicted probability of the correct category; This is the class weight parameter, used to balance the importance of different classes. For tumor segmentation tasks, it is set to [value]. This is an adjustment parameter that controls the degree to which the weights of easily distinguishable samples are reduced; it is set to 2.0. It is logarithmic loss. The focus loss function is an improved version of the cross-entropy loss, through... This method reduces the weight of easily distinguishable samples, causing the model to focus more on difficult-to-distinguish samples. This is particularly meaningful for early tumor screening, as early-stage tumors are often difficult-to-distinguish samples in the training data. For example, in CT images of lung nodules, early nodules with a diameter <3mm account for only a very small proportion of the entire image, and using focus loss can significantly improve the detection sensitivity for such small nodules.

[0091] The disease classification module 3 is connected to the image segmentation module 2. It receives the segmentation results from the image segmentation module 2, determines the disease type based on the segmentation results, and generates suspected lesion detection results. The disease classification module 3 adopts a ResNet network structure, extracting image features and outputting image classification results through a network composed of convolutional layers and pooling layers.

[0092] In this embodiment, the ResNet network structure includes multiple residual blocks, each consisting of two 3×3 convolutional layers and short-circuit connections, effectively solving the gradient vanishing problem in deep network training. The network input is the segmentation result from the image segmentation module, and the output is the corresponding disease type and confidence level. Preferably, the network depth is 50 or 101, and the output layer uses the softmax function for multi-class classification, providing corresponding predicted probabilities for different types of tumors (such as benign, malignant, early-stage cancer, etc.). For example, in breast cancer screening, the system can classify lesions into categories such as normal, benign (e.g., fibroadenoma, cyst), suspicious (requiring further examination), and malignant (e.g., invasive ductal carcinoma); in lung cancer screening, it can classify them into categories such as normal, inflammatory lesions, benign nodules, and malignant nodules.

[0093] Risk assessment module 4 is connected to image segmentation module 2, and is used to receive the segmentation results from image segmentation module 2, perform risk assessment on suspicious lesions, and generate risk probability output. Figure 5 As shown, the risk assessment module 4 includes a labeling unit 41, a risk model construction unit 42, and an assessment unit 43.

[0094] The annotation unit 41 is used to label regions of suspicious lesions in the segmentation results and generate the gold standard. The risk model construction unit 42 is used to construct X-GBoost and residual models to calculate the risk probability of suspicious lesions. The evaluation unit 43 is used to evaluate the performance of the risk prediction model using ROC curves.

[0095] In this embodiment, the X-GBoost model is an ensemble learning algorithm based on gradient boosting trees, which can effectively handle complex feature relationships in medical images. This model receives image features (such as shape, texture, and intensity) extracted from segmentation results as input and outputs the risk probability of the lesion. The residual model is a deep learning-based risk assessment model that addresses the problems in training deep networks through residual connections. The combination of these two models improves the accuracy and robustness of risk assessment. Risk assessment results are represented by probability values ​​between 0 and 1, with higher values ​​indicating a higher risk of malignancy. Preferably, a risk value greater than 0.7 is considered high risk, requiring further clinical examination; a risk value between 0.3 and 0.7 is considered medium risk, with regular follow-up recommended; and a risk value less than 0.3 is considered low risk.

[0096] For example, in lung nodule risk assessment, the system comprehensively considers factors such as nodule size, shape, margin characteristics, density, and growth rate. A nodule with an 8mm diameter, irregular margins, and heterogeneous density may be assessed as high risk (risk value 0.85), while a nodule with a 4mm diameter, smooth margins, and homogeneous density may be assessed as low risk (risk value 0.15). This risk stratification helps clinicians develop personalized follow-up and intervention strategies.

[0097] The system of the present invention may also include an image acquisition unit, an edge computing unit, and a cloud storage / cloud service unit, forming a more complete early tumor screening solution.

[0098] The image acquisition unit is used to acquire medical images of patients through medical imaging equipment. In this embodiment, the image acquisition unit can be a conventional medical imaging scanning device, such as an X-ray machine, a computed tomography (CT) scanner, an ultrasound scanner, and an endoscope. These devices are connected to the edge computing unit via the Internet or a local area network to achieve image acquisition and transmission. For example, in a lung cancer screening project at a primary care hospital, a low-dose CT scanner can be used as the image acquisition unit to perform screening CT examinations on high-risk groups (such as long-term smokers), and then the acquired images are transmitted to the edge computing unit for analysis.

[0099] The edge computing unit is connected to the image acquisition unit to receive medical images acquired by the image acquisition unit and perform local image processing and analysis. In this embodiment, the core of the edge computing unit consists of a high-performance computer, a high-resolution display, and corresponding hardware and software, and has the functions of medical image acquisition, transmission, image analysis, and result evaluation. The edge computing unit is mainly responsible for image acquisition, intelligent processing and evaluation, and local storage of image data and results, and connects to a cloud storage / cloud service unit via the Internet. In a specific application scenario, the edge computing unit can be installed in the radiology department of a hospital, equipped with a workstation with GPU acceleration capabilities (such as a workstation equipped with an NVIDIA RTX 3090 graphics card), which can complete the processing and preliminary analysis of a single CT sequence in a short time (about 15-20 seconds), providing real-time auxiliary diagnosis for radiologists.

[0100] The cloud storage / cloud service unit is connected to the edge computing unit to provide cloud-based data storage, artificial intelligence computing, and remote diagnosis and treatment services. In this embodiment, the cloud service unit is provided by a cloud server and is responsible for cloud-based data storage, artificial intelligence computing, and remote diagnosis and treatment services. The cloud service unit includes a tumor detection model classification result evaluation module, an abnormal tumor risk assessment module, a manual evaluation and intervention module, and a remote expert consultation and remote assistance module. By analyzing the obtained classification results, it locates abnormal areas, determines the proportion of abnormal areas, and analyzes abnormal areas to obtain the lesion characteristics of abnormal tumors and assess the risk of benign or malignant tumors. This architecture is particularly suitable for regional medical collaboration networks. For example, in county and township medical systems, county-level hospitals can deploy edge computing units, while experts from municipal or provincial hospitals can provide remote consultation services through the cloud service unit, effectively solving the problem of insufficient professional radiologists in primary healthcare institutions.

[0101] Preferably, the system of this invention adopts a data-driven design, establishing an image acquisition-edge computing-cloud service unit for early tumor screening via the Internet or local area network, and assisting clinical decision-making through a data-driven approach. The system establishes a neural network framework based on multi-center data, combining it with the image acquisition unit for image grading and annotation to complete tumor screening. The edge computing unit and cloud storage / service unit train images using multi-class deep learning based on a multi-layer network structure, including but not limited to convolutional neural networks, generative adversarial deep neural networks, residual deep neural networks, Transformer models, and deep learning networks, and employs multi-layer training optimization for intelligent evaluation and assisted screening. In practical deployments, such as nationwide lung cancer screening projects, lung image data from different regions, ethnicities, and CT equipment can be collected, and a more generalized lung nodule detection model can be trained through federated learning while protecting data privacy.

[0102] To verify the effectiveness of this invention, we conducted clinical validation tests on the system. The test dataset included 1000 chest CT images, 800 mammogram images, and 600 cervical TCT images, encompassing both benign and malignant cases confirmed by pathology. These data came from routine clinical examinations at three tertiary-level hospitals, covering patients of different ages, genders, and pathological types, ensuring data diversity and representativeness.

[0103] During the model training phase, following the method described in this invention, 3 / 5 of the data from both the benign lesion group and the malignant lesion group are used to train the suspicious lesion detection model, 3 / 10 of the data are used to train the risk prediction model, and the remaining 1 / 10 of the data is used to generate a test set for testing overall performance. For example, for the lung nodule detection task, the benign group contains 600 CT scans (360 used to train the detection model, 180 used to train the risk prediction model, and the remaining 60 used for testing), and the malignant group contains 400 CT scans (240 used to train the detection model, 120 used to train the risk prediction model, and the remaining 40 used for testing). This stratified sampling strategy ensures the consistency of the distribution of training and testing data, improving the reliability of model evaluation.

[0104] Test results show that, compared with traditional deep learning-based tumor detection systems, this invention demonstrates significant advantages in the early detection of small tumors:

[0105] 1. The detection rate of small targets is improved by 42.3%, especially for early lesions with a diameter of less than 5 mm. For example, in the detection of lung nodules, the traditional method has a detection rate of only 35% for tiny nodules with a diameter of 2-3 mm, while the system of this invention can achieve 78%, which has important clinical significance for the early screening of lung cancer.

[0106] 2. The segmentation accuracy (Dice coefficient) was improved by 18.6%, and the average error in boundary localization was reduced to 2.1 pixels. This means that the system can more accurately locate tumor boundaries, providing a more reliable basis for subsequent treatment planning and surgical navigation. In breast tumor segmentation, the Dice coefficient of this system reached 0.87, significantly higher than the 0.73 of traditional methods.

[0107] 3. The false positive rate was reduced by 43.7%, significantly decreasing unnecessary further examinations. In lung cancer screening, traditional methods produce an average of 3.5 false positive results per CT sequence, while this system produces only 1.97, which can significantly reduce the psychological burden on patients and the load on the medical system.

[0108] 4. The processing time for a single CT sequence is controlled within 17.5 seconds, meeting the needs of real-time clinical processing. Even for high-resolution chest CT sequences (512×512×350), this system can complete the analysis within 20 seconds on a standard workstation, enabling doctors to obtain immediate feedback during the examination.

[0109] The test data mentioned above comes from the clinical validation trial of the system of this invention conducted at Yunnan Cancer Hospital (the Third Affiliated Hospital of Kunming Medical University). The test dataset contains medical imaging data from 2400 pathologically confirmed cases, including 1000 chest CT images, 800 mammogram X-ray images, and 600 cervical TCT examination images. The baseline system for the comparative experiment uses a commonly used clinical tumor detection system based on the U-Net architecture. All tests were performed on a standard workstation equipped with an NVIDIA RTX 3090 graphics card, and the test results were independently reviewed and confirmed by three radiology experts with the rank of associate chief physician or above.

[0110] These results fully validate the effectiveness and advancement of the image segmentation method based on differential geometry theory in early tumor screening. In particular, the manifold feature extraction unit, curvature analysis enhancement unit, and geodesic flow attention fusion unit of this invention work synergistically to significantly improve the system's ability to detect early, small lesions, providing strong technical support for early tumor detection and intervention.

[0111] In summary, the AI-assisted early tumor screening image analysis system provided by this invention organically combines differential geometry theory with deep learning technology to construct a complete medical image analysis framework. The system's innovations in manifold representation, curvature analysis, and geodesic flow attention mechanism give it significant advantages in the detection of early, small tumors, providing a new technical approach to improve the accuracy and efficiency of early tumor screening.

[0112] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. An artificial intelligence-assisted image analysis system for early tumor screening, characterized in that, include: The data preprocessing module is used to convert medical images from chest CT, mammography X-ray, and cervical TCT examinations into DICOM format. It performs grayscale normalization, contrast adjustment, and size scaling on different types of medical images, and scales the medical images to make them suitable for subsequent model analysis. An image segmentation module, connected to the data preprocessing module, is used to receive a normalized medical image sent by the data preprocessing module and segment suspicious disease areas in the normalized medical image based on differential geometric feature analysis. The image segmentation module includes a manifold feature extraction unit, a curvature analysis enhancement unit, and a geodesic flow attention fusion unit. The manifold feature extraction unit is used to: map the normalized medical image into a two-dimensional manifold representation; calculate the metric tensor and curvature information on the two-dimensional manifold representation; construct a multi-scale feature pyramid to extract geometric features at different scales; and generate a feature map containing local geometric structure information. The curvature analysis enhancement unit is used to: identify potential abnormal regions based on the Gaussian curvature and average curvature of the feature map; construct a curvature feature map to highlight potential lesion regions; dynamically determine the enhancement threshold according to the distribution of the curvature feature map; perform adaptive geometric feature enhancement on the identified potential lesion regions; and output the enhanced feature map. For regions exceeding the threshold, a geometric feature enhancement function is applied to enhance their feature representation. ， in, The enhanced eigenvalue at point p; It is the value of the original feature map at point p; and These represent the absolute values ​​of the Gaussian curvature and the mean curvature deviating from the mean, respectively. It is the strength enhancement coefficient. For point Gaussian curvature at that point; For point The average curvature at that point, It is the mean of the Gaussian curvature of all points in the image; It is the standard deviation of Gaussian curvature; It is the mean of the average curvature of all points in the image; It is the standard deviation of the mean curvature; The geodesic flow attention fusion unit is used to: define geodesic flows on the characteristic manifold and characterize the feature evolution process, wherein the geodesic flows are described by the following partial differential equations: ， in, Representation of features Regarding time The partial derivatives describe the rate of change of a feature over time; It is the initial feature; Based on Riemannian metric The gradient operator; It is a feature With initial features Geodetic distance between them; Key feature regions are identified based on geodesic flow evolution patterns. A three-layer feature fusion network is constructed, including a feature extraction layer, a feature selection layer, and a fusion output layer. The feature extraction layer includes residual connection units, convolutional feature units, and geometric feature units. The feature selection layer includes a geodesic flow-guided attention control unit and a dual-path processing unit, wherein the dual-path processing unit includes a shallow path for preserving detailed information and a deep path for extracting semantic information. The fusion output layer includes a feature integration unit, a segmentation prediction unit, and a geometric refinement unit. Multi-level features are integrated through a geodesic flow-guided attention mechanism to generate the final segmentation result. The geodesic flow attention fusion unit is optimized using a focus loss function during training: ， in, The value of the focus loss function; It is the model's predicted probability of the correct category; This is the category weight parameter, set to 0.25; This is an adjustment parameter; set it to 2.

0. It uses logarithmic loss; and employs a multi-stage training strategy to first train each sub-network, and then performs end-to-end fine-tuning. The disease classification module, connected to the image segmentation module, is used to receive the segmentation results from the image segmentation module, determine the disease type from the segmentation results, and generate suspicious lesion detection results. A risk assessment module, connected to the image segmentation module, is used to receive the segmentation results from the image segmentation module. The risk assessment module includes: a labeling unit, used to label the regions of suspicious lesions in the segmentation results and generate a gold standard; a risk model construction unit, used to construct X-GBoost and residual models to calculate the risk probability of the suspicious lesions; and an evaluation unit, used to evaluate the performance of the risk prediction model using ROC curves.

2. The artificial intelligence-assisted early tumor screening image analysis system according to claim 1, characterized in that, In the manifold feature extraction unit, medical images Viewed as a two-dimensional manifold embedded in a higher-dimensional space Through mapping function Establish a correspondence; the Gaussian curvature and mean curvature They are respectively: ， ， in, and It is a point Principal curvature at the point; It is an image At point The gradient vector at that point.

3. The artificial intelligence-assisted early tumor screening image analysis system according to claim 1, characterized in that, In the curvature analysis enhancement unit, the dynamic determination of the enhancement threshold is achieved in the following way: ， ， in, The dynamic threshold for Gaussian curvature; It refers to adjusting parameters; The dynamic threshold for the average curvature; It refers to adjusting parameters.

4. The artificial intelligence-assisted early tumor screening image analysis system according to claim 1, characterized in that, In the geodesic flow attention fusion unit, the geodesic distance is defined as: ， in, Represent two points on the characteristic manifold and Geodetic distance between them; This means taking all connections. and curve The minimum value in; The parameter is The curve, ; Is the curve in The tangent vector at the point; It is a point on the curve The Riemannian metric tensor at that location.

5. The artificial intelligence-assisted early tumor screening image analysis system according to claim 1, characterized in that, The disease classification module adopts a ResNet network structure, which extracts image features and outputs image classification results through a network composed of convolutional layers and pooling layers.

6. The artificial intelligence-assisted early tumor screening image analysis system according to claim 1, characterized in that, The system further includes: an image acquisition unit for acquiring medical images of patients through medical imaging equipment; an edge computing unit connected to the image acquisition unit for receiving the medical images acquired by the image acquisition unit and performing local image processing and analysis; and a cloud storage / cloud service unit connected to the edge computing unit for providing cloud-based data storage, artificial intelligence computing, and remote diagnosis and treatment services.

Images