An artificial intelligence-based tumor recognition auxiliary system
By using an AI-based tumor identification assistance system, which dynamically adjusts CT scan parameters and fuses multimodal features, the problem of tumor identification relying on manual diagnosis is solved, achieving efficient and accurate tumor identification and radiation dose optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HANGZHOU DEEP LOVE TECH CO LTD
- Filing Date
- 2025-06-26
- Publication Date
- 2026-06-12
AI Technical Summary
In existing technologies, tumor identification on CT images mainly relies on manual diagnosis, which involves subjectivity and the risk of misdiagnosis. Furthermore, the lack of personalized CT scan parameters leads to unnecessary increases in radiation dose and low diagnostic efficiency.
An AI-based tumor identification assistance system is adopted. The system acquires medical record information and tumor location through an information acquisition module, dynamically adjusts CT scan parameters, and uses multimodal feature fusion and deep learning models for tumor identification. This reduces repetitive work and diagnostic discrepancies, and optimizes CT parameters to reduce radiation dose.
It enables efficient visualization of tumor boundaries and internal structures, reduces radiation dose, improves diagnostic accuracy, reduces false positives, minimizes diagnostic discrepancies among physicians, and enhances the automation and personalization of tumor identification.
Smart Images
Figure CN120707544B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and in particular to an artificial intelligence-based tumor identification assistance system. Background Technology
[0002] Currently, imaging examinations are one of the most direct and effective methods for tumor identification. On CT images, tumors clinically manifest as several irregularly shaped tissue regions of varying sizes within the organ parenchyma, with indistinct boundaries. Because early-stage tumors often present with no obvious symptoms or atypical imaging findings, they are difficult to detect and diagnose, leading to most cases being diagnosed at an advanced stage, missing the optimal treatment window. Therefore, early diagnosis and identification of tumors are crucial for cancer patients. In early tumor diagnosis, multi-slice spiral CT, through reconstruction technology, can clearly display lesion characteristics in transverse, sagittal, and coronal planes. In intermediate-stage diagnosis, spiral CT combined with surface masking and multiplanar reconstruction can clearly show the tumor location, internal structure, marginal features, blood supply, degree of invasion of surrounding tissues, and changes in surrounding tissues, achieving high diagnostic accuracy. Therefore, CT images have become an important reference for tumor diagnosis and identification.
[0003] However, the current process of diagnosing tumors through medical images mainly relies on manual methods and experienced doctors. This process is subjective, and errors can occur due to limitations in doctors' cognitive abilities or fatigue, leading to misdiagnosis. These shortcomings highlight the importance of using effective medical image analysis technology to improve the accuracy of disease diagnosis.
[0004] Therefore, there is a need to provide an artificial intelligence-based tumor identification assistance system for the automatic analysis of medical images. Summary of the Invention
[0005] This invention provides an artificial intelligence-based tumor identification assistance system, comprising: an information acquisition module for acquiring patient medical record information and tumor identification location information; an image acquisition module for determining the patient's initial CT parameters based on the patient's medical record information and tumor identification location information, acquiring the patient's initial CT image based on the patient's initial CT parameters, extracting image features from the patient's initial CT image, determining multiple sets of target CT parameters for the patient based on the image features of the patient's initial CT image, and acquiring multiple target CT images for the patient based on the multiple sets of target CT parameters; and a identification assistance module for establishing and training a tumor identification model, performing tumor identification based on the patient's multiple target CT images using the tumor identification model, and generating tumor identification assistance information.
[0006] Furthermore, the image acquisition module determines the patient's initial CT parameters based on the patient's medical record information and tumor identification location information, including: for each tumor identification location, acquiring the medical record factors and historical imaging data that affect the CT imaging of the tumor identification location, wherein the historical imaging data of the tumor identification location includes the medical record information, CT images, and CT parameters of multiple historical patients; dividing the multiple historical patients into multiple historical patient groups based on the medical record factors that affect the CT imaging of the tumor identification location, and determining the initial CT parameters of each historical patient group; querying multiple historical patient groups based on the patient's tumor identification location information; querying the target historical patient group from the multiple historical patient groups based on the medical record factors that affect the CT imaging of the patient's tumor identification location and the patient's medical record information; and determining the patient's initial CT parameters based on the initial CT parameters of the target historical patient group.
[0007] Furthermore, based on the influence of CT imaging on tumor identification location on medical records, multiple historical patients are divided into multiple historical patient groups, and the initial CT parameters for each historical patient group are determined. This includes: determining the medical record feature vector of each historical patient based on the influence of CT imaging on tumor identification location on medical records and the medical record information of the historical patients; dividing multiple historical patients into multiple historical patient groups using a clustering algorithm based on the medical record feature vectors of multiple historical patients; establishing a fitness function; for each historical patient group, determining the fitness of the CT parameters of the historical patients based on the fitness function and the CT images of each historical patient included in the historical patient group; and determining the initial CT parameters of the historical patient group using a genetic algorithm based on the CT parameters of each historical patient included in the historical patient group and the fitness of the CT parameters.
[0008] Furthermore, a fitness function is established, including: determining multiple CT image quality assessment factors; determining a radiation dose calculation function, wherein the independent variables of the radiation dose calculation function include CT parameters; and establishing a fitness function based on the multiple CT image quality assessment factors and the radiation dose calculation function.
[0009] Furthermore, image features of the patient's initial CT images are extracted, including: for each tumor identification location, determining multiple tumor identification image feature factors associated with the tumor identification location; and extracting image features of the patient's initial CT images based on the multiple tumor identification image feature factors associated with the patient's tumor identification location.
[0010] Furthermore, the method identifies multiple tumor identification image feature factors associated with tumor identification locations, including: identifying multiple tumor identification image feature factors; acquiring initial CT images of multiple historical users in a historical patient group corresponding to the tumor identification location; for each tumor identification image feature factor, calculating the difference value of the tumor identification image feature factor at the tumor identification location based on the initial CT images of multiple historical users; and determining multiple tumor identification image feature factors associated with the tumor identification location based on the difference value of each tumor identification image feature factor at the tumor identification location.
[0011] Furthermore, based on the image features of the patient's initial CT images, multiple sets of target CT parameters for the patient are determined, including: acquiring initial CT images and multiple sets of target CT parameters of multiple historical users corresponding to the tumor identification location in the historical patient group; determining target historical users based on the image features of the patient's initial CT images and the image features of the initial CT images of historical users of the patient's tumor identification location; and determining multiple sets of target CT parameters for the patient based on the target historical users.
[0012] Furthermore, the tumor recognition model includes multiple feature extraction units, a multimodal feature fusion unit, a clinical information embedding unit, and a classification decision unit. Each feature extraction unit corresponds to a set of target CT parameters. The feature extraction unit is used to extract feature vectors from the input target CT image. The multimodal feature fusion unit is used to fuse the feature vectors extracted by multiple feature extraction units and output the fused feature vector. The clinical information embedding unit is used to concatenate the patient's medical record information with the fused feature vector to generate a comprehensive feature vector. The classification decision unit is used to output the tumor recognition result based on the comprehensive feature vector.
[0013] Furthermore, the feature extraction unit includes a first convolutional layer, a first activation function, and multiple dense residual attention blocks, wherein the dense residual attention blocks include an input layer, a dense connection layer, a pointwise convolutional layer, an attention mechanism layer, a residual connection layer, and an output layer.
[0014] Furthermore, the loss function used to train the tumor recognition model includes classification loss and orthogonal constraint loss, wherein the orthogonal constraint loss is calculated based on the feature vector of the input target CT image extracted by multiple feature extraction units.
[0015] Compared to existing technologies, the artificial intelligence-based tumor identification assistance system provided in this specification has at least the following beneficial effects:
[0016] 1. By dynamically adjusting subsequent CT scan parameters (such as contrast agent dose, scan phase, slice thickness, etc.) based on the image features of the initial CT images (such as tumor location, size, density, etc.), multiple sets of target CT images are obtained, reducing unnecessary radiation dose, avoiding the uniform use of high-dose scans, improving the visualization of tumor boundaries and internal structures, and providing higher-quality data for subsequent models. The tumor identification model adopts a multi-branch network or fusion network architecture, which can process multiple sets of target CT images simultaneously, extract multimodal features, avoid the limitations of single-modal information, improve classification performance, and automatically complete CT parameter optimization, image feature extraction, and preliminary tumor identification. Doctors only need to review the results, reducing repetitive work. Through standardized processes and deep learning models, diagnostic differences between different doctors are reduced.
[0017] 2. Based on the patient's tumor location and medical history factors (such as age, weight, and medical history), target patient groups are selected from historical data, and their optimized CT parameters are inherited to avoid a "one-size-fits-all" scanning scheme. The quality of CT parameters is quantitatively evaluated through a fitness function to balance image quality (such as signal-to-noise ratio and contrast) and radiation dose. The optimal CT parameters are globally searched through a genetic algorithm to avoid local optima. Through personalized parameter optimization, high-quality images can be obtained on the first scan, reducing repeated scans caused by inappropriate parameters.
[0018] 3. An independent feature extraction unit is assigned to each set of target CT parameters to extract tumor features under different target CT parameters, avoiding information loss. A clinical information embedding unit is used to stitch patient medical history information (such as age and medical history) with image features, improving the model's ability to identify tumors and reducing false positives (e.g., misdiagnosing benign nodules as tumors in elderly patients). Orthogonal constraint loss forces the output feature vectors of multiple feature extraction units to be orthogonal, reducing redundant information and improving the complementarity of multimodal features. Attached Figure Description
[0019] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein:
[0020] Figure 1 This is a block diagram of an artificial intelligence-based tumor identification assistance system shown in one embodiment of this application;
[0021] Figure 2 This is a structural diagram of a tumor recognition model shown in one embodiment of this application. Detailed Implementation
[0022] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below.
[0023] Figure 1 This is a block diagram of an artificial intelligence-based tumor recognition assistance system shown in one embodiment of this application, such as... Figure 1 As shown, an artificial intelligence-based tumor recognition assistance system may include an information acquisition module, an image acquisition module, and a recognition assistance module.
[0024] The information acquisition module can be used to acquire the patient's medical records and tumor location information.
[0025] Specifically, a patient's medical record information may include basic information such as gender, age, BMI, blood routine test results, and biochemical indicators, as well as medical history. The information acquisition module can retrieve the patient's medical record information from the hospital's data system.
[0026] Tumor location information refers to the location that needs to be identified through a CT scan. The information acquisition module can obtain the patient's tumor location information from the doctor's terminal.
[0027] In some embodiments, the image acquisition module determines the patient's initial CT parameters based on the patient's medical record information and tumor identification location information, including:
[0028] For each tumor identification location, the medical history factors and historical imaging data of the CT imaging of the tumor identification location are obtained. The historical imaging data of the tumor identification location includes medical history information, CT images and CT parameters of multiple historical patients. Based on the medical history factors of the CT imaging of the tumor identification location, the multiple historical patients are divided into multiple historical patient groups, and the initial CT parameters of each historical patient group are determined. The CT imaging locations of multiple historical patients are all the tumor identification locations.
[0029] Based on the patient's tumor identification location information, query multiple historical patient groups. For example, historical patient groups 1-3 correspond to tumor identification location A, and historical patient groups 4-7 correspond to tumor identification location B. Based on the patient's tumor identification location information, the patient's tumor identification location is A, so historical patient groups 1-3 are used as the multiple historical patient groups obtained from the query.
[0030] Based on the factors affecting the patient's medical history and the patient's medical history information, the target historical patient group is queried from multiple historical patient groups. For example, based on the factors affecting the patient's medical history and the patient's medical history information, the patient's medical history feature vector is determined, and the Euclidean distance between the patient's medical history feature vector and the medical history feature vector of historical patients sampled from the historical patient group is calculated. The historical patient group to which the historical patients whose Euclidean distance is less than the Euclidean distance threshold belong is taken as the target historical patient group.
[0031] Based on the initial CT parameters of the target historical patient group, the initial CT parameters of the patient are determined. These CT parameters may include parameters such as tube voltage, tube current, and slice thickness. For example, the initial CT parameters of the target historical patient group can be used as the initial CT parameters of the patient.
[0032] Specifically, factors influencing CT imaging of tumor location can include those related to sex, age, BMI, and disease. For example, sex hormones (such as estrogen and testosterone) may affect the growth and metabolism of certain tumors, thus influencing tumor characteristics (such as density and enhancement patterns) in CT imaging. With age, tissue density may change (such as decreased bone density and increased adipose tissue), affecting the measurement of CT values (Hounsfield Unit, HU). High BMI patients have increased fat and soft tissue thickness, leading to increased X-ray attenuation, which may reduce the image signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR). Surgery, radiotherapy, and chemotherapy may alter the anatomical structure or biological characteristics of tumors. Comorbidities (such as diabetes and cardiovascular disease) may affect a patient's hemodynamics and metabolic status, indirectly affecting CT imaging.
[0033] CT imaging factors that influence medical records can be determined manually (e.g., by experts, doctors, etc.) to identify tumor locations.
[0034] In some embodiments, based on CT imaging influence on medical history factors related to tumor identification location, multiple historical patients are divided into multiple historical patient groups, and initial CT parameters for each historical patient group are determined, including:
[0035] Based on CT imaging factors affecting tumor identification location and medical record information of historical patients, determine the medical record feature vector of historical patients.
[0036] Using a clustering algorithm, multiple historical patients are divided into multiple historical patient groups based on the medical record feature vectors of multiple historical patients.
[0037] Establish the fitness function;
[0038] For each historical patient group, the fitness of the CT parameters of the historical patients is determined based on the fitness function and the CT images of each historical patient included in the historical patient group. The initial CT parameters of the historical patient group are determined by a genetic algorithm based on the CT parameters of each historical patient included in the historical patient group and the fitness of the CT parameters.
[0039] Specifically, determining the medical record feature vector of historical patients based on CT imaging factors affecting tumor identification location and historical patient medical record information may include the following steps:
[0040] S11. Medical Record Information Extraction: Extract information from historical patients' electronic medical records that corresponds to factors influencing CT imaging findings related to tumor location, such as: Demographic characteristics: gender, age, BMI; Laboratory tests: Complete blood count (hemoglobin, platelets), biochemical indicators (liver function, kidney function); Disease characteristics: Tumor type, stage, location, size; Previous treatments: Surgical history, radiotherapy history, chemotherapy history.
[0041] S12. Feature Quantization: Convert non-numerical features (such as gender and tumor type) into numerical codes. For example: Gender: Male = 1, Female = 0. Tumor type: Lung cancer = 1, Liver cancer = 2, Breast cancer = 3.
[0042] S13. Feature Vector Construction: Combine all quantified features into a multi-dimensional vector, for example: Feature Vector = [Gender, Age, BMI, Hemoglobin, Tumor Type, Tumor Size]. Example: Male, 55 years old, BMI=25, Hemoglobin=130g / L, Lung Cancer, Tumor Size=3cm, the corresponding medical record feature vector is: [1, 55, 25, 130, 1, 3].
[0043] The K-means clustering algorithm is used to cluster multiple historical patients into multiple historical patient groups. The specific steps include:
[0044] Initialization: Randomly select K historical patients as center points (K is the preset number of groups).
[0045] Assignment: The Euclidean distance between the feature vectors of patients not selected as center points and the feature vectors of the center points can be calculated, and each patient's feature vector can be assigned to the group containing the center point with the smallest Euclidean distance.
[0046] Update centroids: Recalculate the centroids (mean eigenvectors) for each group.
[0047] Iteration: Repeat the assignment and update steps until the center point no longer changes or the maximum number of iterations is reached.
[0048] In some embodiments, establishing a fitness function includes:
[0049] Determine multiple CT image quality assessment factors, such as spatial resolution, noise level, contrast, and artifact level;
[0050] Determine the radiation dose calculation function, where the independent variables of the radiation dose calculation function include CT parameters;
[0051] A fitness function is established based on multiple CT image quality assessment factors and radiation dose calculation functions.
[0052] For example, the radiation dose calculation function can be:
[0053]
[0054] in, This is the dose-length product, reflecting the total radiation dose in a single scan. For tube current, For tube voltage, Where is the pitch, and a, b, c, d are empirical coefficients related to the CT equipment model and scanning protocol. For the scan length, 'a' ranges from 0.001 to 0.1, representing the contribution of tube current to the dose; 'b' ranges from 0.1 to 10, used to adjust the baseline dose level, and its value varies depending on the equipment model and scanning protocol; 'c' ranges from 1.5 to 3.0, reflecting the influence of tube voltage on the dose; and 'd' ranges from 0.5 to 2.0, reflecting the influence of pitch on the dose.
[0055] For example, the fitness function can be:
[0056]
[0057] in, For the fitness function, , , , and As weight, , , , and Greater than 0, Example =0.2、 =0.1、 =0.2、 =0.2、 =0.3, The normalized score for spatial resolution (0~1) represents the image quality; higher CT image quality corresponds to a higher normalized score for spatial resolution. The normalized contrast score (0~1) represents the contrast ratio; the higher the CT image quality, the higher the normalized contrast score. The noise level is represented by a normalized score (0-1). Higher CT image quality corresponds to a lower normalized noise level score. The score is a normalized score (0-1) for the degree of artifacts. The higher the CT image quality, the lower the normalized score for the degree of artifacts. For the normalized score of radiation dose (0~1), After normalization, we get , The larger the value, the higher the normalized score of the radiation dose.
[0058] Understandably, the fitness function comprehensively evaluates the quality of CT images (spatial resolution, contrast, noise level, and artifact severity) and radiation dose through a weighted summation. Higher normalized scores for spatial resolution and contrast result in larger fitness function values; higher normalized scores for noise, artifacts, and radiation dose result in smaller fitness function values. By utilizing the fitness function, it is possible to minimize radiation dose while ensuring CT image quality meets diagnostic requirements, thereby improving patient safety and comfort.
[0059] In some embodiments, determining the initial CT parameters of a historical patient group based on the CT parameters and fitness of each historical patient included in the historical patient group using a genetic algorithm may include the following steps:
[0060] (1) Initialize the population
[0061] Historical imaging data for tumor identification location includes CT parameters from multiple historical patients as multiple initial individuals.
[0062] (2) Calculate fitness
[0063] For each individual (CT parameters), calculate its fitness value:
[0064] Calculate radiation dose based on CT parameters.
[0065] Based on historical CT images of patients, assess image quality factors (such as spatial resolution, noise level, etc.).
[0066] Substitute the values into the fitness function to calculate the fitness value.
[0067] (3) Select
[0068] Select high-quality individuals based on their fitness values (individuals with high fitness values are more likely to be selected).
[0069] (4) Cross
[0070] Two parent individuals are randomly selected, and offspring individuals are generated through a crossover operation.
[0071] Intersection methods: single-point intersection, multi-point intersection, etc.
[0072] (5) Variation
[0073] To increase population diversity, certain CT parameters of offspring individuals were randomly perturbed.
[0074] (6) Termination Conditions
[0075] Repeat steps (2) to (5) until the termination condition is met: the maximum number of iterations or the fitness value converges.
[0076] (7) Output the optimal solution
[0077] The individual with the highest fitness value is returned as the initial CT parameters for the historical patient group.
[0078] The image acquisition module can be used to determine the patient's initial CT parameters based on the patient's medical record information and tumor identification location information, acquire the patient's initial CT image based on the patient's initial CT parameters, extract the image features of the patient's initial CT image, determine multiple sets of target CT parameters of the patient based on the image features of the patient's initial CT image, and acquire multiple target CT images of the patient based on the multiple sets of target CT parameters of the patient.
[0079] In some embodiments, the image acquisition module extracts image features from the patient's initial CT image, including:
[0080] For each tumor identification location, multiple tumor identification image feature factors associated with the tumor identification location are determined. These multiple tumor identification image feature factors are used to quantify the morphological, texture, metabolic and other characteristics of the tumor. For example, morphological feature factors (e.g., the number of pixels of the tumor in the two-dimensional image, the pixel length of the tumor boundary, etc.), texture feature factors (e.g., gray-level co-occurrence matrix, gray-level run matrix), intensity feature factors (e.g., average gray value, maximum / minimum gray value, skewness, kurtosis, etc.).
[0081] Based on multiple tumor identification image feature factors associated with the patient's tumor identification location, image features of the patient's initial CT image are extracted.
[0082] Specifically, suspected tumor regions are located in the patient's initial CT images using image segmentation or object detection algorithms (such as U-Net, Mask R-CNN, etc.). The grayscale values of the suspected tumor region image are normalized to a fixed range (e.g., [0, 1] or [-1, 1]), and Gaussian filtering or median filtering is used to remove noise, unify the image resolution, and ensure feature consistency. Based on the tumor mask, morphological features of the patient's initial CT images are extracted, and texture and intensity features are extracted based on the pixel values of the patient's initial CT images. In other words, the image features of the patient's initial CT images include morphological features, texture features, and intensity features.
[0083] In some embodiments, the image acquisition module determines multiple tumor identification image feature factors associated with the tumor identification location, including:
[0084] Multiple tumor identification image feature factors can be determined manually (e.g., by doctors or experts).
[0085] Acquire initial CT images of multiple historical users in a historical patient group corresponding to the tumor identification location. The initial CT images of historical users can be CT images acquired with the tumor identification location as the CT scan location and under the CT scan parameters corresponding to the historical patient group.
[0086] For each tumor identification image feature factor, based on the initial CT images of multiple historical users, the difference value of the tumor identification image feature factor at the tumor identification location is calculated. For example, it can be based on the tumor identification image feature factor in each sample.
[0087] Based on the difference value of each tumor identification image feature factor at the tumor identification location, multiple tumor identification image feature factors associated with the tumor identification location are determined. For example, tumor identification image feature factors with a difference value greater than the difference value threshold can be used as multiple tumor identification image feature factors associated with the tumor identification location.
[0088] Specifically, for each historical patient group corresponding to the tumor identification location, the variance of the tumor identification image feature factor values for each historical user within that patient group can be calculated as the difference value of the tumor identification image feature factor across the historical patient group corresponding to the tumor identification location. The mean of these difference values for the tumor identification image feature factor across each historical patient group corresponding to the tumor identification location is then calculated and used as the difference value of the tumor identification image feature factor across the tumor identification location.
[0089] In some embodiments, the image acquisition module determines multiple sets of target CT parameters for the patient based on image features of the patient's initial CT images, including:
[0090] Acquire initial CT images and multiple sets of target CT parameters for multiple historical users in the historical patient group corresponding to the tumor identification location. The multiple sets of target CT parameters for the historical patient group corresponding to the tumor identification location can be determined manually (e.g., by doctors or experts).
[0091] Based on the image features of the patient's initial CT images and the image features of the patient's tumor identification location from the initial CT images of historical users, target historical users are identified, and multiple sets of target CT parameters for the patient are determined based on the target historical users.
[0092] Specifically, the cosine similarity between the image features of the patient's initial CT image and the image features of the initial CT images of historical users in the target historical patient group can be calculated. The historical user with the highest cosine similarity is selected as the target historical user. Multiple sets of target CT parameters of the target historical user are then used as multiple sets of target CT parameters for the patient.
[0093] The identification assistance module can be used to build and train a tumor identification model, which can then identify tumors based on multiple target CT images of the patient and generate tumor identification assistance information.
[0094] Figure 2 This is a structural diagram of a tumor recognition model shown in one embodiment of this application, as follows: Figure 2 As shown, in some embodiments, the tumor recognition model includes multiple feature extraction units, a multimodal feature fusion unit, a clinical information embedding unit, and a classification decision unit. Each feature extraction unit corresponds to a set of target CT parameters. The feature extraction unit is used to extract tumor-related feature vectors from the input target CT image, which may include morphological, texture, and intensity features. The multimodal feature fusion unit is used to fuse the feature vectors extracted by multiple feature extraction units, outputting a fused feature vector, and generating a fused feature vector by directly concatenating multiple feature vectors into a longer vector to integrate tumor information under different parameters. The clinical information embedding unit is used to concatenate the patient's medical record information with the fused feature vector to generate a comprehensive feature vector. The classification decision unit is used to output the tumor recognition result based on the comprehensive feature vector, and the classification decision unit may include a support vector machine.
[0095] In some embodiments, the feature extraction unit includes a first convolutional layer, a first activation function, and a plurality of dense residual attention blocks, wherein the dense residual attention blocks include an input layer, a dense connection layer, a pointwise convolutional layer, an attention mechanism layer, a residual connection layer, and an output layer.
[0096] Specifically, the first convolutional layer performs preliminary feature extraction on the input CT image (e.g., using a 3×3 convolutional kernel) to generate a basic feature map. The first activation function is ReLU, which enhances the model's expressive power. High-level features are further extracted through multiple stacked dense residual attention blocks, which alleviates the gradient vanishing problem and enhances feature expressive power.
[0097] Dense residual attention blocks are modules that combine dense connections, residual connections, and attention mechanisms. Their structure is as follows:
[0098] 1. Input layer:
[0099] It receives the feature map from the previous layer as input.
[0100] 2. Dense Connection Layer:
[0101] Function: By using dense connections, the input of the current layer is concatenated with the outputs of all previous layers, enabling feature reuse and gradient flow.
[0102] accomplish:
[0103] Assuming the current layer is the l-th layer, its input is the concatenation of the outputs of all previous layers: xl=[x0,x1,…,xl−1].
[0104] Dense connections can reduce the number of parameters, alleviate the vanishing gradient problem, and enhance feature propagation.
[0105] Output: The concatenated feature map.
[0106] 3. Pointwise Convolution Layer:
[0107] Function: Performs channel dimensionality reduction or expansion on densely connected feature maps using 1×1 convolutions, reducing computational cost and adjusting the number of channels.
[0108] accomplish:
[0109] Using a 1×1 convolution kernel, each location of the feature map is convolved independently, without changing the spatial dimension, only adjusting the number of channels.
[0110] Output: Feature map after adjusting the number of channels.
[0111] 4. Attention Mechanism Layer:
[0112] Function: Dynamically adjust the importance of different positions or channels in the feature map through attention mechanisms (such as channel attention, spatial attention, or hybrid attention).
[0113] accomplish:
[0114] Channel attention: Channel weights are generated using global average pooling (GAP) and fully connected layers (FC) to weight each channel of the feature map.
[0115] Spatial attention: Use convolutional layers to generate spatial weight maps, which weight each spatial location of the feature map.
[0116] Hybrid Attention: Combining channel attention and spatial attention to generate a more refined weight map.
[0117] Output: Attention-weighted feature map.
[0118] 5. Residual Connection Layer:
[0119] Function: By adding the input feature map to the attention-weighted feature map through residual connection, feature reuse and gradient flow are achieved.
[0120] accomplish:
[0121] The residual connection formula is: y = F(x) + x, where F(x) is the attention-weighted feature map and x is the input feature map.
[0122] Residual connections can alleviate the vanishing gradient problem in deep networks and accelerate model convergence.
[0123] 6. Output: Feature map after residual connection.
[0124] Output layer:
[0125] Output the final feature map of the current dense residual attention block as the input to the next layer.
[0126] In some embodiments, the loss function used to train the tumor recognition model includes a classification loss and an orthogonal constraint loss, wherein the orthogonal constraint loss is calculated based on the feature vector of the input target CT image extracted by multiple feature extraction units.
[0127] For example, the loss function used to train a tumor recognition model is:
[0128]
[0129]
[0130]
[0131]
[0132] in, For the total loss, For classifying losses, For orthogonal constraint loss, These are hyperparameters used to balance the weights of the classification loss and the orthogonality constraint loss. For the number of categories, For real labels, Let be the predicted probability of the c-th class for the k-th training sample. The orthogonal constraint loss for the k-th training sample is... For batch size, The total number of feature extraction units. The feature vector of the target CT image extracted by the i-th feature extraction unit is the input feature vector of the target CT image. The feature vector of the target CT image extracted by the j-th feature extraction unit. for and The inner product, For feature vectors L2 norm, For feature vectors L2 norm, It is the square of the L2 norm, used to measure the cosine similarity between two feature vectors.
[0133] As can be understood, in the above formula, the classification loss is used to optimize the model's classification performance, while the orthogonality constraint loss is used to ensure diversity among the feature vectors extracted by different feature extraction units. The classification loss is calculated using cross-entropy loss and measures the difference between the model's predicted probability and the true label. The orthogonality constraint loss is calculated using the squared cosine similarity of the feature vectors and is used to constrain the orthogonality between the feature vectors extracted by different feature extraction units. The squared cosine similarity measures the directional consistency between two feature vectors (0 when orthogonal, 1 when parallel). By minimizing the orthogonality constraint loss, the feature vectors extracted by different feature extraction units are forced to be orthogonal, avoiding feature redundancy. By combining the classification loss and the orthogonality constraint loss, the model can extract more discriminative and diverse features while optimizing classification performance, thereby improving the robustness and accuracy of tumor recognition tasks.
[0134] Finally, it should be understood that the embodiments described in this specification are merely illustrative of the principles of the embodiments described herein. Other variations may also fall within the scope of this specification. Therefore, alternative configurations of the embodiments described herein are intended to be illustrative rather than limiting, and should be considered consistent with the teachings of this specification. Accordingly, the embodiments described herein are not limited to those explicitly introduced and described herein.
Claims
1. An artificial intelligence-based tumor identification assistance system, characterized in that, include: The information acquisition module is used to acquire the patient's medical record information and tumor identification location information; The image acquisition module is used to determine the patient's initial CT parameters based on the patient's medical record information and tumor identification location information, acquire the patient's initial CT image based on the patient's initial CT parameters, extract the image features of the patient's initial CT image, determine multiple sets of target CT parameters of the patient based on the image features of the patient's initial CT image, and acquire multiple target CT images of the patient based on the multiple sets of target CT parameters of the patient. The tumor identification assistance module is used to establish and train a tumor identification model, which identifies tumors based on multiple target CT images of the patient and generates tumor identification assistance information. The image acquisition module determines the patient's initial CT parameters based on the patient's medical record information and tumor location information, including: For each tumor identification location, the medical history factors and historical imaging data of the CT imaging at the tumor identification location are acquired. The historical imaging data of the tumor identification location includes medical history information, CT images and CT parameters of multiple historical patients. Based on the medical history factors of the CT imaging at the tumor identification location, the multiple historical patients are divided into multiple historical patient groups, and the initial CT parameters of each historical patient group are determined. Based on the patient's tumor location information, query multiple historical patient groups; Based on the CT imaging factors affecting the patient's tumor identification location and the patient's medical record information, the target historical patient group is queried from multiple historical patient groups. Determine the initial CT parameters of the patients based on the initial CT parameters of the target historical patient group; Based on the influence of CT imaging on tumor location on medical history factors, multiple historical patients were divided into multiple historical patient groups, and initial CT parameters for each historical patient group were determined, including: Based on CT imaging factors affecting tumor identification location and medical record information of historical patients, determine the medical record feature vector of historical patients. Using a clustering algorithm, multiple historical patients are divided into multiple historical patient groups based on the medical record feature vectors of multiple historical patients. Establish the fitness function; For each historical patient group, the fitness of the CT parameters of the historical patients is determined based on the fitness function and the CT images of each historical patient included in the historical patient group. The initial CT parameters of the historical patient group are determined by a genetic algorithm based on the CT parameters of each historical patient included in the historical patient group and the fitness of the CT parameters. Establish the fitness function, including: Determine multiple CT image quality assessment factors; Determine the radiation dose calculation function, where the independent variables of the radiation dose calculation function include CT parameters; A fitness function is established based on multiple CT image quality assessment factors and radiation dose calculation functions. The image acquisition module determines multiple sets of target CT parameters for the patient based on the image features of the patient's initial CT images, including: Acquire initial CT images and multiple sets of target CT parameters for multiple historical users in the historical patient group corresponding to the tumor identification location; Based on the image features of the patient's initial CT images and the image features of the patient's tumor identification location from the initial CT images of historical users, target historical users are identified, and based on the target historical users, multiple sets of target CT parameters for the patient are determined. The tumor identification model includes multiple feature extraction units, a multimodal feature fusion unit, a clinical information embedding unit, and a classification decision unit. Each feature extraction unit corresponds to a set of target CT parameters. The feature extraction unit is used to extract feature vectors from the input target CT image. The multimodal feature fusion unit is used to fuse the feature vectors extracted by multiple feature extraction units and output the fused feature vector. The clinical information embedding unit is used to concatenate the patient's medical record information with the fused feature vector to generate a comprehensive feature vector. The classification decision unit is used to output the tumor identification result based on the comprehensive feature vector.
2. The tumor identification assistance system based on artificial intelligence according to claim 1, characterized in that, The image acquisition module extracts image features from the patient's initial CT images, including: For each tumor identification location, determine multiple tumor identification image feature factors associated with the tumor identification location; Based on multiple tumor identification image feature factors associated with the patient's tumor identification location, image features of the patient's initial CT image are extracted.
3. The tumor identification assistance system based on artificial intelligence according to claim 2, characterized in that, The image acquisition module determines multiple tumor identification image feature factors associated with the tumor identification location, including: Identify multiple tumor recognition image feature factors; Acquire initial CT images of multiple historical users in the historical patient group corresponding to the tumor identification location; For each tumor identification image feature factor, the difference value of the tumor identification image feature factor at the tumor identification location is calculated based on the initial CT images of multiple historical users; Based on the difference values of each tumor identification image feature factor at the tumor identification location, multiple tumor identification image feature factors associated with the tumor identification location are determined.
4. The tumor identification assistance system based on artificial intelligence according to claim 1, characterized in that, The feature extraction unit includes a first convolutional layer, a first activation function, and multiple dense residual attention blocks, wherein the dense residual attention blocks include an input layer, a dense connection layer, a pointwise convolutional layer, an attention mechanism layer, a residual connection layer, and an output layer.
5. The tumor identification assistance system based on artificial intelligence according to claim 4, characterized in that, The loss function used to train the tumor recognition model includes classification loss and orthogonal constraint loss, wherein the orthogonal constraint loss is calculated based on the feature vector of the input target CT image extracted by multiple feature extraction units.