Deep learning-based CT image analysis and tumor prediction method and system

Abstract

The invention discloses a CT image analysis and tumor prediction method and system based on deep learning. The method comprises the following steps: collecting CT image data and annotation information of a tumor patient to obtain a high-quality CT image data set; a neural structure search technology and a composite model expansion method are combined to establish a second-generation efficient convolutional neural network tumor classification model; adding a coordinate attention module and a hard smooth activation function to the model to obtain an improved second-generation efficient convolutional neural network tumor classification model; carrying out training and optimization on the basis of the model; and inputting the CT image data of the patient into the optimized second-generation efficient convolutional neural network tumor classification model to obtain a classification result, and predicting the existence, category or malignancy degree of the tumor according to the classification result. The system comprises a data acquisition module, a model establishment module, a model improvement module, an optimization module and a result output module. According to the invention, the analysis efficiency of the CT image and the accuracy of tumor diagnosis can be obviously improved.

Description

Technical Field

[0001] The present invention relates to the field of tumor prediction technology, and in particular to a CT image analysis and tumor prediction method and system based on deep learning. Background Art

[0002] In recent years, with the development of medical imaging technology, CT (Computed Tomography) has become one of the important tools for tumor detection. CT images are widely used to detect and analyze the shape, location and size of tumors with their high resolution and precise anatomical structure information. However, traditional medical image analysis methods rely on the professional experience and naked eye judgment of doctors. This diagnostic method is highly subjective and easily affected by factors such as visual fatigue, which may lead to misdiagnosis or missed diagnosis. With the increasing number of brain tumor patients year by year, this traditional method obviously can no longer meet the needs of modern medicine for efficient and accurate diagnosis.

[0003] In the field of medical image analysis, computer-aided diagnosis (CAD) systems are gradually becoming an important assistant for doctors due to their high efficiency and consistency. CAD systems can significantly shorten diagnosis time, improve diagnostic accuracy, and reduce errors in manual operations through automated processing and analysis of CT images. However, traditional CAD systems mostly rely on manual feature extraction and classification algorithms based on classical machine learning, which have the following major defects: low feature extraction efficiency. Traditional methods require manual feature design, which is not only time-consuming, but the extracted features may not fully reflect the complex characteristics of tumors; limited classification accuracy. Traditional classification algorithms (such as support vector machines or random forests) are prone to overfitting or classification errors when processing high-dimensional complex data; insufficient data utilization. Traditional methods are difficult to fully utilize the potential information of large-scale medical image datasets.

[0004] In recent years, the application of deep learning technology, especially convolutional neural network (CNN), in the field of medical image processing has attracted widespread attention. Deep learning models can automatically extract multi-level features from medical images and effectively capture complex tumor morphology and texture information through deep neural network structures. Compared with traditional methods, they have the following significant advantages: automatic feature extraction, CNN can automatically learn the best features from images, reducing dependence on manual design; improved classification performance, deep learning models usually perform better than traditional machine learning models in medical image classification, and can achieve higher classification accuracy; multi-task processing capabilities, by designing multi-task learning models, deep learning can simultaneously complete tumor classification, segmentation and feature analysis tasks.

[0005] Although deep learning has shown powerful capabilities in medical image analysis, the following technical problems still exist: high data annotation costs. High-quality medical image datasets require precise manual annotation, especially for the segmentation and classification of tumor areas. The annotation work is time-consuming and requires high professional knowledge; data is scarce. Medical image datasets are usually small in scale, which limits the generalization ability of deep learning models; model complexity. Deep learning models have complex structures and high requirements for hardware equipment and computing resources, which increases the difficulty of actual deployment.

[0006] The existing technologies in CT image analysis and tumor prediction have the following main deficiencies: low diagnostic efficiency. Traditional manual analysis of CT images takes a long time and cannot meet the needs of rapid diagnosis; low diagnostic accuracy. Due to the lack of automated feature extraction and analysis tools, the diagnostic results of existing systems are easily affected by subjective factors; lack of dedicated models. Existing deep learning models are not optimized for CT image characteristics and cannot fully utilize the spatial and texture information in CT data; insufficient artifact processing. Metal artifacts or motion artifacts may exist in CT images, affecting the identification and classification of tumor areas; high difficulty in model application. The deployment and application of existing deep learning models in actual medical scenarios are constrained by hardware limitations and difficulties in data acquisition. Summary of the invention

[0007] In view of this, an object of an embodiment of the present invention is to provide a CT image analysis and tumor prediction method and system based on deep learning, which can significantly improve the analysis efficiency of CT images and the accuracy of tumor diagnosis.

[0008] The embodiment of the present invention is achieved as follows:

[0009] A CT image analysis and tumor prediction method based on deep learning, comprising:

[0010] CT image data and annotation information of tumor patients are collected to obtain a medical image data set, and the medical image data set is preprocessed to obtain a high-quality CT image data set.

[0011] Combining neural structure search technology and composite model expansion method, a second-generation efficient convolutional neural network tumor classification model was established.

[0012] A coordinate attention module and a hard smoothing activation function are added to the second-generation efficient convolutional neural network tumor classification model to obtain an improved second-generation efficient convolutional neural network tumor classification model.

[0013] Based on the improved second-generation efficient convolutional neural network tumor classification model, training and optimization are performed.

[0014] The patient's CT image data is input into the optimized second-generation efficient convolutional neural network tumor classification model to obtain the classification results, and the existence, category or malignancy of the tumor is predicted based on the classification results.

[0015] In a preferred embodiment of the present invention, in the above-mentioned deep learning-based CT image analysis and tumor prediction method, the collecting of CT image data and annotation information of tumor patients to obtain a medical image data set, and preprocessing of the medical image data set to obtain a high-quality CT image data set include:

[0016] Collect CT image data of tumor patients with tumor types in designated locations and pathological diagnosis reports of tumor patients to obtain the specific location and pathological type of CT images and lesions.

[0017] A medical image annotation tool is used to annotate the CT image, a corresponding annotation mask is generated, and the annotation information is matched with the corresponding CT image to obtain a medical image data set.

[0018] Adaptive threshold segmentation and morphological operations are used to remove metal artifacts or motion artifacts in the CT image.

[0019] The different CT images were normalized to a fixed resolution using linear interpolation.

[0020] The CT image is smoothed and denoised using a non-local means algorithm.

[0021] The Laplacian operator is used to enhance the boundary of the lesion area in the CT image.

[0022] The processed CT image data is converted into a format for a deep learning model and stored according to tumor type, stage or organ classification to obtain a high-quality CT image dataset.

[0023] Its technical effects are: removing metal artifacts and motion artifacts through adaptive threshold segmentation and morphological operations, effectively reducing the interference of artifacts on image quality, using non-local mean algorithm to smooth and reduce noise of CT images to effectively suppress image noise, and using Laplace operator to enhance the boundary of lesion area to make the lesion features more prominent. Linear interpolation is used to standardize CT images to a fixed resolution to ensure consistency of all images, solve the problem of resolution differences caused by different image acquisition equipment and technical parameters, convert the processed CT image data into a format that can be used by the deep learning model, and store it according to tumor type, stage or organ classification, so as to facilitate efficient model call and training.

[0024] In a preferred embodiment of the present invention, in the above-mentioned deep learning-based CT image analysis and tumor prediction method, the combination of neural structure search technology and composite model expansion method to establish a second-generation efficient convolutional neural network tumor classification model includes:

[0025] The search space is set for the second-generation efficient convolutional neural network tumor classification model. The shallow layer uses a fused moving bottleneck convolution module to quickly extract low-level features of the input CT image data. The deep layer uses a deep separable moving bottleneck convolution module and uses depthwise separable convolution and attention mechanism for feature extraction.

[0026] The fused moving bottleneck convolution module includes a common convolution kernel with a width of 3 pixels and a height of 3 pixels.

[0027] The depth separation moving bottleneck convolution module includes a dimensionality increase convolution kernel with a width of 1 pixel and a height of 1 pixel, a depth separable convolution kernel with a width of 3 pixels and a height of 3 pixels, and a dimensionality reduction convolution kernel with a width of 1 pixel and a height of 1 pixel.

[0028] Taking classification accuracy and computational cost as optimization goals, the objective function maximize Accuracy(Net(d,w,r))-λ·Cost(Net) was set to automatically find the optimal network structure, where d is the depth, w is the width, r is the input resolution, Cost(Net) is the resource cost, and λ is the weight for balancing classification performance and computational cost. The optimal configuration of the second-generation efficient convolutional neural network tumor classification model was obtained.

[0029] The composite scaling factor φ is set to make the second generation efficient convolutional neural network tumor classification model proportionally expand d=α in depth d, width w and input resolution r. φ , w=β φ , r = γ φ , establish the constraint relationship α·β of the parameters of the second generation efficient convolutional neural network tumor classification model 2 γ 2 ≈2,α≥1,β≥1,γ≥1, and the depth d, width w, and input resolution r of the neural network are uniformly scaled through the composite model expansion method to optimize the second-generation efficient convolutional neural network tumor classification model.

[0030] The structure of the second-generation efficient convolutional neural network tumor classification model after compound scaling adjustment is: in, Indicates module stacking, FL i is the calculation function of the i-th convolution module, X is the input feature map, H i is the height of the input image, W i is the width of the input image, Ci is the number of channels in each layer.

[0031] Its technical effects are as follows: the fusion mobile bottleneck convolution module efficiently extracts low-level features of the input CT image, retains the basic features and reduces the computational complexity. The deep separation mobile bottleneck convolution module uses the dimension-raising convolution kernel to enhance the feature dimension. The deep separable convolution kernel extracts fine-grained features and reduces redundant features through the dimension-reducing convolution kernel. Combined with the attention mechanism, the model's ability to focus on key areas of lesions is enhanced, thereby improving classification accuracy. By setting the search space, taking classification accuracy and computational cost as the target optimization function, and automatically searching for the optimal network structure, the time cost of manual debugging is significantly reduced, ensuring that the model achieves optimal performance in tumor classification tasks and finding the best balance between performance and resource consumption. By stacking the optimized convolution modules, each layer gradually strengthens the feature extraction and expression capabilities, especially when processing complex CT images, showing stronger classification capabilities.

[0032] In a preferred embodiment of the present invention, in the above-mentioned deep learning-based CT image analysis and tumor prediction method, the coordinate attention module and the hard smoothing activation function are added to the second-generation efficient convolutional neural network tumor classification model to obtain an improved second-generation efficient convolutional neural network tumor classification model, which includes:

[0033] A hard smoothing activation function is added to the fused mobile bottleneck convolution module.

[0034] An attention module and a hard smoothing activation function are added to the depth-separated mobile bottleneck convolution module.

[0035] The attention module globally pools the feature map input to the depth separation mobile bottleneck convolution module, extracts features, concatenates and maps them to a low-dimensional space, and then returns to the original space, and enhances the spatial dimension of the input feature map in a weighted manner.

[0036] The hard smoothing activation function translates, truncates and normalizes the feature map input into the fused moving bottleneck convolution module or the deep separation moving bottleneck convolution module, and outputs the complex texture and morphological features of the tumor.

[0037] The technical effect is as follows: the coordinate attention module globally pools the input feature map, maps high-dimensional features to low-dimensional space and then returns to the original space, and weights the feature map in the spatial dimension, so that the model can focus on the tumor area, improve the recognition ability of the key area of ​​the tumor, and accurately distinguish the tumor from the surrounding tissue by enhancing the spatial features. After adding the hard smoothing activation function, the model is more accurate in extracting low-level features (such as the rough shape and size of the tumor) at the shallow level, while reducing the interference of background noise on feature extraction; it extracts texture, density and morphological features more efficiently at the deep level, and further enhances the expression ability of key areas by combining the attention mechanism; for complex pathological manifestations of tumors (such as unclear boundaries, internal structural heterogeneity, etc.), the improved model can better capture small but important features. The coordinate attention module enhances the model's discrimination in the spatial dimension by learning the interaction between global and local features, more accurately locates the boundary area of ​​the tumor, and has a stronger recognition ability for tumors in complex anatomical structures (such as tumors closely connected to adjacent organs).

[0038] In a preferred embodiment of the present invention, in the above-mentioned deep learning-based CT image analysis and tumor prediction method, the adding of a hard smoothing activation function to the fused moving bottleneck convolution module comprises:

[0039] The shallow input feature map of the fused mobile bottleneck convolution module is X1.

[0040] The common convolution kernel is combined with the hard smoothing activation function to obtain the fused output Z of the fused mobile bottleneck convolution module. low =Hard-Swish(W1*X1+b1), where W1 is the weight of the shallow input feature map and b1 is the bias term.

[0041] The formula of the hard smooth activation function is x is the element value of the shallow input feature map, ReLU6(x) is the activation function, and the input is truncated within the interval [0,6]. The formula is ReLU6(x)=min(max(0,x),6), which means that the element value of the shallow input feature map is shifted so that some negative values ​​participate in the activation.

[0042] Its technical effect is: by translating, truncating and normalizing the element values ​​of the shallow input feature map, the hard smooth activation function limits the input to the interval [0,6], ensuring that the smaller negative value part of the input feature map can still participate in the activation, avoiding the problem that some features cannot be activated in the conventional ReLU function; by translating and truncating, some negative values ​​in the feature map can participate in the subsequent learning process, which helps to improve the model's learning of detailed features, especially when the tumor boundary and structure are blurred, it can effectively capture richer tumor features. After the shallow input feature map is processed by the ordinary convolution kernel, the fusion output generated by the hard smooth activation function can significantly enhance the texture and morphological characteristics of the tumor; especially in CT images, tumors often show complex boundaries and morphology. The hard smooth activation function enhances the perception of complex tumor textures by better retaining the details near the boundaries. This is particularly important for the identification and classification of malignant tumors.

[0043] In a preferred embodiment of the present invention, in the above-mentioned CT image analysis and tumor prediction method based on deep learning, the adding of an attention module to the deep separation moving bottleneck convolution module comprises:

[0044] The depth input feature map of the deep separation moving bottleneck convolution module is

[0045] For the feature map Perform global pooling in the vertical and horizontal directions respectively to obtain the horizontal encoding features and vertical encoding features Among them, c is the channel index, w is the horizontal position of the feature map, and h is the vertical position of the feature map.

[0046] The horizontal encoding feature Y w and the vertical encoding feature Y h Through the cascade operation fusion, the compact feature representation Z = ReLU (W1 · [Y h ; Y w ]+b1), where the compact feature representation W1 is the dimensionality reduction weight, b1 is the bias parameter, and ReLU(·) is the nonlinear activation function.

[0047] Map the compact feature representation Z back to the horizontal direction to obtain the horizontal mapping feature g w (w,c)=Sigmoid(W w ·Z+b w ), Map back to the vertical direction to obtain the vertical mapping feature g h (h,c)=Sigmoid(Wh ·Z+b h ),

[0048] Use attention weights to weight features X CA (h,w,c)=X in (h,w,c)·g h (h,c)·g w (w,c), get the output features of the attention module

[0049] The technical effect is: the feature map is globally pooled in the horizontal and vertical directions respectively to obtain the encoding features in the horizontal and vertical directions, which not only retains the long-distance dependency of the image in space, but also enhances the model's detailed understanding of the tumor structure by combining the features in the horizontal and vertical directions, especially when the tumor morphology is more complex or there are subtle differences. By connecting the horizontal encoding features and the vertical encoding features in series, and then obtaining a compact feature representation through a nonlinear activation function and dimensionality reduction weights, unnecessary redundant information is effectively reduced, while retaining important high-dimensional features, so that the model can focus more on analyzing the key features of the tumor during subsequent processing, improve classification accuracy, and reduce the consumption of computing resources. The compact feature representation after dimensionality reduction is mapped back to the horizontal and vertical directions, and the horizontal mapping features and vertical mapping features obtained effectively restore the structural information of the tumor in space, and more accurately locate the position and morphological characteristics of the tumor, especially in complex CT images, where the tumor may be mixed with normal tissue or located at the edge of the organ. The attention mechanism can help the model make more accurate judgments.

[0050] In a preferred embodiment of the present invention, in the above-mentioned deep learning-based CT image analysis and tumor prediction method, the training and optimization based on the improved second-generation efficient convolutional neural network tumor classification model includes:

[0051] The high-quality CT image dataset is divided into a training dataset and a validation set.

[0052] The training data set is input into the improved second-generation efficient convolutional neural network tumor classification model for forward propagation to obtain a prediction output

[0053] Calculate the cross entropy loss function Where y is the predicted output The corresponding label, C is the number of categories, y i is the i-th label of the true label, is the predicted probability for the i-th class.

[0054] Perform back propagation and update the model parameters by calculating the gradient of the loss function in, is the updated parameter, η1 is the learning rate, is the cross entropy loss function About weight W l gradient.

[0055] Adjust weights through the Adam optimization algorithm Among them, W t is the weight after the tth iteration, m t is the momentum of the gradient, v t is the average value of the square of the gradient, ò is a constant, and η2 is the learning rate.

[0056] The validation set is used to validate the trained second-generation efficient convolutional neural network tumor classification model.

[0057] Its technical effect is that through the combination of reasonable data division, loss function calculation, optimization algorithm and verification mechanism, it ensures that the deep learning model can achieve higher accuracy, robustness and generalization ability when processing complex CT image data. At the same time, by using the Adam optimization algorithm and cross entropy loss function, the model training process is not only more efficient and stable, but also avoids overfitting problems, improving the accuracy and reliability of tumor prediction.

[0058] In a preferred embodiment of the present invention, in the above-mentioned deep learning-based CT image analysis and tumor prediction method, the use of the verification set to verify the trained second-generation efficient convolutional neural network tumor classification model includes:

[0059] Load the final parameters of the second-generation efficient convolutional neural network tumor classification model after training.

[0060] The validation set is input into the improved second-generation efficient convolutional neural network tumor classification model for forward propagation to obtain a prediction output

[0061] Output the prediction Compare with the true label of the sample, calculate the accuracy, precision, and recall rate. If the performance of the validation set does not meet expectations, it is determined that the hyperparameters of the model need to be adjusted or the network architecture needs to be modified, and the parameters need to be adjusted to obtain the model.

[0062] Its technical effect is: by inputting the validation set into the trained model and calculating indicators such as accuracy, precision, and recall, a comprehensive evaluation of the model performance is achieved, which can effectively avoid the overfitting and underfitting problems of the model, and provide feedback for adjusting hyperparameters or modifying the network architecture. It not only improves the accuracy, stability, and robustness of the model, but also provides an important basis for the continuous optimization of the model.

[0063] In a preferred embodiment of the present invention, in the above-mentioned deep learning-based CT image analysis and tumor prediction method, the patient's CT image data is input into the optimized second-generation efficient convolutional neural network tumor classification model to obtain a classification result, and the existence, category or malignancy of the tumor is predicted according to the classification result, including:

[0064] The patient's CT image data is preprocessed.

[0065] The pre-processed CT image data is input into the optimized second-generation efficient convolutional neural network tumor classification model to perform shallow feature extraction, deep feature extraction, attention enhancement and classification prediction to obtain the prediction results representing the probability distribution of each category by the model.

[0066] Select the category with the highest probability As the final prediction, if , then it is judged that the tumor exists.

[0067] According to category c pred The specific meaning of the output prediction result is used, and the usage probability p is used cpred as the confidence level of the prediction.

[0068] Its technical effect is: by inputting the patient's CT image data into the optimized second-generation efficient convolutional neural network tumor classification model, combining shallow and deep feature extraction, attention enhancement and probability prediction mechanism, it can accurately predict the existence, category and malignancy of the tumor. It not only provides accurate tumor classification results, but also provides confidence for doctors, improves the reliability and efficiency of diagnosis, and has strong clinical application value.

[0069] A CT image analysis and tumor prediction system based on deep learning, comprising:

[0070] The data acquisition module is used to collect CT image data and annotation information of tumor patients to obtain a medical image data set, and pre-process the medical image data set to obtain a high-quality CT image data set.

[0071] The model building module is used to combine neural structure search technology and composite model expansion method to establish a second-generation efficient convolutional neural network tumor classification model.

[0072] The model improvement module is used to add a coordinate attention module to the second-generation efficient convolutional neural network tumor classification model to obtain an improved second-generation efficient convolutional neural network tumor classification model.

[0073] The optimization module is used to perform training and optimization based on the improved second-generation efficient convolutional neural network tumor classification model.

[0074] The result output module is used to input the patient's CT image data into the optimized second-generation efficient convolutional neural network tumor classification model to obtain the classification results, and predict the existence, category or malignancy of the tumor based on the classification results.

[0075] The beneficial effects of the embodiments of the present invention are:

[0076] Through a sophisticated preprocessing process, the present invention can effectively remove artifacts in CT images, such as metal artifacts and motion artifacts, and ensure the image quality of the input deep learning model. Through adaptive threshold segmentation, morphological operations, smooth noise reduction processing, boundary enhancement and other technologies, the significance of the lesion area is effectively improved, which is helpful for the subsequent feature extraction and classification process.

[0077] Combining neural structure search technology and composite model expansion method, the present invention constructs a second-generation efficient convolutional neural network (CNN) tumor classification model, using a fusion mobile bottleneck convolution module and a depth separation mobile bottleneck convolution module, which effectively improves the accuracy of feature extraction. The design also takes into account the optimization of computing resources, adopts an automated optimization search space strategy, ensures the balance between depth, width and resolution, and achieves the optimal balance between accuracy and computing efficiency in the classification model.

[0078] The present invention further adds a coordinate attention module and a hard smooth activation function to the deep learning model. The coordinate attention module can enhance the spatial dimension of the input feature map and improve the model's ability to capture fine-grained features, thereby improving the precise positioning and classification of tumors. The introduction of the hard smooth activation function helps to enhance the model's ability to extract complex texture and morphological features of tumors, thereby improving the stability and robustness of the prediction.

[0079] During the training process, the present invention divides the high-quality CT image data set into a training set and a validation set, and uses the cross entropy loss function and the Adam optimization algorithm for training. The model weights are adjusted by back propagation, the performance of the model is verified using the validation set, and evaluation indicators such as accuracy, precision, and recall are calculated. If the performance of the model on the validation set is not ideal, the hyperparameters of the model can be adjusted in time to avoid overfitting problems and ensure the generalization ability of the model.

[0080] The optimized convolutional neural network model can extract shallow and deep features from the patient's CT images and accurately predict the presence, type and malignancy of the tumor. The prediction results not only include whether the tumor exists, but also output probability as confidence, which improves the interpretability of tumor prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0082] Figure 1 This is a flow chart of the CT image analysis and tumor prediction method based on deep learning of the present invention. DETAILED DESCRIPTION

[0083] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0084] Please refer to Figure 1 The first embodiment of the present invention provides a CT image analysis and tumor prediction method based on deep learning, which includes: collecting CT image data and annotation information of tumor patients to obtain a medical image data set, preprocessing the medical image data set to obtain a high-quality CT image data set; combining neural structure search technology and composite model expansion method to establish a second-generation efficient convolutional neural network tumor classification model; adding a coordinate attention module and a hard smoothing activation function to the second-generation efficient convolutional neural network tumor classification model to obtain an improved second-generation efficient convolutional neural network tumor classification model; training and optimizing on the basis of the improved second-generation efficient convolutional neural network tumor classification model; inputting the patient's CT image data into the optimized second-generation efficient convolutional neural network tumor classification model to obtain a classification result, and predicting the existence, category or malignancy of the tumor according to the classification result.

[0085] In a preferred embodiment of the present invention, in the above-mentioned deep learning-based CT image analysis and tumor prediction method, the CT image data and annotation information of tumor patients are collected to obtain a medical image data set, and the medical image data set is preprocessed to obtain a high-quality CT image data set, including: collecting CT image data of tumor patients with tumor types in specified locations and pathological diagnosis reports of tumor patients to obtain the specific location and pathological type of the CT image and lesions; using a medical image annotation tool (such as 3D Slicer or ITK-SNAP) to annotate the CT image, generate a corresponding annotation mask, match the annotation information with the corresponding CT image, and obtain a medical image data set; using adaptive threshold segmentation and morphological operations to remove metal artifacts or motion artifacts in the CT image; using linear interpolation to standardize different CT images to a fixed resolution; using a non-local mean algorithm to smooth and reduce noise on the CT image; using a Laplace operator to enhance the boundary of the lesion area in the CT image; converting the processed CT image data into a format for a deep learning model, and storing it according to tumor type, stage or organ classification to obtain a high-quality CT image data set.

[0086] Its technical effects are: removing metal artifacts and motion artifacts through adaptive threshold segmentation and morphological operations, effectively reducing the interference of artifacts on image quality, using non-local mean algorithm to smooth and reduce noise of CT images to effectively suppress image noise, and using Laplace operator to enhance the boundary of lesion area to make the lesion features more prominent. Linear interpolation is used to standardize CT images to a fixed resolution to ensure consistency of all images, solve the problem of resolution differences caused by different image acquisition equipment and technical parameters, convert the processed CT image data into a format that can be used by the deep learning model, and store it according to tumor type, stage or organ classification, so as to facilitate efficient model call and training.

[0087] In a preferred embodiment of the present invention, in the above-mentioned CT image analysis and tumor prediction method based on deep learning, the combination of neural structure search technology and composite model expansion method to establish a second-generation efficient convolutional neural network tumor classification model includes: setting a search space for the second-generation efficient convolutional neural network tumor classification model, using a fused moving bottleneck convolution module in the shallow layer to quickly extract low-level features from the input CT image data, and using a deep separation moving bottleneck convolution module in the deep layer to extract features using deep separable convolution and attention mechanism; the fused moving bottleneck convolution module includes a common convolution kernel with a width of 3 pixels and a height of 3 pixels; the deep separation moving bottleneck convolution module includes a dimensionality-raising convolution kernel with a width of 1 pixel and a height of 1 pixel, a deep separable convolution kernel with a width of 3 pixels and a height of 3 pixels, and a dimensionality-reducing convolution kernel with a width of 1 pixel and a height of 1 pixel; taking classification accuracy and computational cost as optimization goals, setting an objective function maximize for automatically finding the optimal network structure Accuracy(Net(d,w,r))-λ·Cost(Net), where d is the depth, w is the width, r is the input resolution, Cost(Net) is the resource cost, and λ is the weight for balancing classification performance and computational cost, to obtain the optimal configuration of the second-generation efficient convolutional neural network tumor classification model; set the composite scaling factor φ so that the second-generation efficient convolutional neural network tumor classification model is proportionally and synchronously expanded in depth d, width w and input resolution r by d=α φ , w=β φ , r = γ φ , establish the constraint relationship α·β of the parameters of the second generation efficient convolutional neural network tumor classification model 2 γ 2 ≈2,α≥1,β≥1,γ≥1, the depth d, width w and input resolution r of the neural network are uniformly scaled by the composite model expansion method to optimize the second-generation efficient convolutional neural network tumor classification model; where α is the depth expansion coefficient, which controls the number of network layers d and expands the network depth to capture more hierarchical features, β is the width expansion coefficient, which controls the number of channels per layer w and enhances the single-layer feature extraction capability, and γ is the resolution expansion coefficient, which controls the size r of the input image and improves the resolution of the input feature map; the square of the width β and the resolution γ are closely related to the convolution operation, and the computational complexity of the convolution is the product of the number of channels and the area of ​​the feature map; the total scaling amount is ≈2, ensuring that the expansion of depth, width and resolution remains balanced under the doubling resource constraint; the structure of the second-generation efficient convolutional neural network tumor classification model after composite scaling adjustment is: in, Indicates module stacking, FL i is the calculation function of the i-th convolution module, X is the input feature map, H i is the height of the input image, Wi is the width of the input image, C i is the number of channels in each layer.

[0088] Its technical effects are as follows: the fusion mobile bottleneck convolution module efficiently extracts low-level features of the input CT image, retains the basic features and reduces the computational complexity. The deep separation mobile bottleneck convolution module uses the dimension-raising convolution kernel to enhance the feature dimension. The deep separable convolution kernel extracts fine-grained features and reduces redundant features through the dimension-reducing convolution kernel. Combined with the attention mechanism, the model's ability to focus on key areas of lesions is enhanced, thereby improving classification accuracy. By setting the search space, taking classification accuracy and computational cost as the target optimization function, and automatically searching for the optimal network structure, the time cost of manual debugging is significantly reduced, ensuring that the model achieves optimal performance in tumor classification tasks and finding the best balance between performance and resource consumption. By stacking the optimized convolution modules, each layer gradually strengthens the feature extraction and expression capabilities, especially when processing complex CT images, showing stronger classification capabilities.

[0089] In a preferred embodiment of the present invention, in the above-mentioned deep learning-based CT image analysis and tumor prediction method, the coordinate attention module and the hard smooth activation function are added to the second-generation efficient convolutional neural network tumor classification model to obtain an improved second-generation efficient convolutional neural network tumor classification model, including: adding a hard smooth activation function to the fused moving bottleneck convolution module; adding an attention module and a hard smooth activation function to the deep separation moving bottleneck convolution module; the attention module globally pools the feature map input to the deep separation moving bottleneck convolution module, extracts features, concatenates and maps them to a low-dimensional space, and then returns to the original space, and enhances the spatial dimension of the input feature map by weighted means; the hard smooth activation function translates, truncates and normalizes the feature map input to the fused moving bottleneck convolution module or the deep separation moving bottleneck convolution module, and outputs the complex texture and morphological features of the tumor.

[0090] The technical effect is as follows: the coordinate attention module globally pools the input feature map, maps high-dimensional features to low-dimensional space and then returns to the original space, and weights the feature map in the spatial dimension, so that the model can focus on the tumor area, improve the recognition ability of the key area of ​​the tumor, and accurately distinguish the tumor from the surrounding tissue by enhancing the spatial features. After adding the hard smoothing activation function, the model is more accurate in extracting low-level features (such as the rough shape and size of the tumor) at the shallow level, while reducing the interference of background noise on feature extraction; it extracts texture, density and morphological features more efficiently at the deep level, and further enhances the expression ability of key areas by combining the attention mechanism; for complex pathological manifestations of tumors (such as unclear boundaries, internal structural heterogeneity, etc.), the improved model can better capture small but important features. The coordinate attention module enhances the model's discrimination in the spatial dimension by learning the interaction between global and local features, more accurately locates the boundary area of ​​the tumor, and has a stronger recognition ability for tumors in complex anatomical structures (such as tumors closely connected to adjacent organs).

[0091] In a preferred embodiment of the present invention, in the above-mentioned deep learning-based CT image analysis and tumor prediction method, the adding of a hard smoothing activation function to the fused moving bottleneck convolution module comprises: the shallow input feature map of the fused moving bottleneck convolution module is X1; the ordinary convolution kernel is combined with the hard smoothing activation function to obtain the fused output Z of the fused moving bottleneck convolution module. low =Hard-Swish(W1*X1+b1), where W1 is the weight of the shallow input feature map and b1 is the bias term; the formula of the hard smooth activation function is x is the element value of the shallow input feature map, ReLU6(x) is the activation function, and the input is truncated within the interval [0,6]. The formula is ReLU6(x)=min(max(0,x),6), which means that the element value of the shallow input feature map is shifted so that some negative values ​​participate in the activation.

[0092] Its technical effect is: by translating, truncating and normalizing the element values ​​of the shallow input feature map, the hard smooth activation function limits the input to the interval [0,6], ensuring that the smaller negative value part of the input feature map can still participate in the activation, avoiding the problem that some features cannot be activated in the conventional ReLU function; by translating and truncating, some negative values ​​in the feature map can participate in the subsequent learning process, which helps to improve the model's learning of detailed features, especially when the tumor boundary and structure are blurred, it can effectively capture richer tumor features. After the shallow input feature map is processed by the ordinary convolution kernel, the fusion output generated by the hard smooth activation function can significantly enhance the texture and morphological characteristics of the tumor; especially in CT images, tumors often show complex boundaries and morphology. The hard smooth activation function enhances the perception of complex tumor textures by better retaining the details near the boundaries. This is particularly important for the identification and classification of malignant tumors.

[0093] In a preferred embodiment of the present invention, in the above-mentioned CT image analysis and tumor prediction method based on deep learning, the adding of the attention module to the deep separation moving bottleneck convolution module includes: the deep input feature map of the deep separation moving bottleneck convolution module is For the feature map Perform global pooling in the vertical and horizontal directions respectively to obtain the horizontal encoding features and vertical encoding features Where c is the channel index, w is the horizontal position of the feature map, and h is the vertical position of the feature map; the horizontal encoding feature Y w and the vertical encoding feature Y h Through the cascade operation fusion, the compact feature representation Z = ReLU (W1 · [Y h ; Y w ]+b1), where the compact feature representation W1 is the dimensionality reduction weight, b1 is the bias parameter, and ReLU(·) is the nonlinear activation function; the compact feature representation Z is mapped back to the horizontal direction to obtain the horizontal mapping feature g w (w,c)=Sigmoid(W w ·Z+b w ), Map back to the vertical direction to obtain the vertical mapping feature g h (h,c)=Sigmoid(W h ·Z+b h ), Use attention weights to weight features X CA (h,w,c)=X in (h,w,c)·gh (h,c)·g w (w,c), get the output features of the attention module

[0094] The technical effect is: the feature map is globally pooled in the horizontal and vertical directions respectively to obtain the encoding features in the horizontal and vertical directions, which not only retains the long-distance dependency of the image in space, but also enhances the model's detailed understanding of the tumor structure by combining the features in the horizontal and vertical directions, especially when the tumor morphology is more complex or there are subtle differences. By connecting the horizontal encoding features and the vertical encoding features in series, and then obtaining a compact feature representation through a nonlinear activation function and dimensionality reduction weights, unnecessary redundant information is effectively reduced, while retaining important high-dimensional features, so that the model can focus more on analyzing the key features of the tumor during subsequent processing, improve classification accuracy, and reduce the consumption of computing resources. The compact feature representation after dimensionality reduction is mapped back to the horizontal and vertical directions, and the horizontal mapping features and vertical mapping features obtained effectively restore the structural information of the tumor in space, and more accurately locate the position and morphological characteristics of the tumor, especially in complex CT images, where the tumor may be mixed with normal tissue or located at the edge of the organ. The attention mechanism can help the model make more accurate judgments.

[0095] In a preferred embodiment of the present invention, in the above-mentioned deep learning-based CT image analysis and tumor prediction method, the training and optimization based on the improved second-generation efficient convolutional neural network tumor classification model includes: dividing the high-quality CT image data set into a training data set and a validation set; inputting the training data set into the improved second-generation efficient convolutional neural network tumor classification model, performing forward propagation, and obtaining a prediction output Calculate the cross entropy loss function Where y is the predicted output The corresponding label, C is the number of categories, y i is the i-th label of the true label, To predict the probability of the i-th class, perform back propagation and update the model parameters by calculating the gradient of the loss function. in, is the updated parameter, η1 is the learning rate, is the cross entropy loss function About weight W l The gradient of ; adjust the weights through the Adam optimization algorithm Among them, W t is the weight after the tth iteration, m t is the momentum of the gradient, v tis the average value of the square of the gradient, ò is a constant, and η2 is the learning rate; the validation set is used to validate the trained second-generation efficient convolutional neural network tumor classification model.

[0096] Its technical effect is that through the combination of reasonable data division, loss function calculation, optimization algorithm and verification mechanism, it ensures that the deep learning model can achieve higher accuracy, robustness and generalization ability when processing complex CT image data. At the same time, by using the Adam optimization algorithm and cross entropy loss function, the model training process is not only more efficient and stable, but also avoids overfitting problems, improving the accuracy and reliability of tumor prediction.

[0097] In a preferred embodiment of the present invention, in the above-mentioned deep learning-based CT image analysis and tumor prediction method, the use of the verification set to verify the trained second-generation efficient convolutional neural network tumor classification model includes: loading the final parameters of the trained second-generation efficient convolutional neural network tumor classification model; inputting the verification set into the improved second-generation efficient convolutional neural network tumor classification model, performing forward propagation, and obtaining a prediction output Output the prediction Compare with the true label of the sample, calculate the accuracy, precision, and recall rate. If the performance of the validation set does not meet expectations, it is determined that the hyperparameters of the model need to be adjusted or the network architecture needs to be modified, and the parameters need to be adjusted to obtain the model.

[0098] Specifically, the accuracy evaluation formula is: in, is an indicator function, which is 1 when the prediction result is consistent with the true label, otherwise it is 0; the evaluation formula for accuracy is Among them, TP is a true positive example, FP is a false positive example; the evaluation formula of recall rate is Among them, FN is a false negative example; the evaluation index also includes F1-score, and the evaluation formula is

[0099] Adjust the model's hyperparameters, including learning rate, batch size, convolution kernel size, etc., and modify the network architecture, including adding or reducing convolution layers, pooling layers, etc.

[0100] The generalization ability of the model is verified by calculating the error of the model on the validation set. If the model performs well on the training set but poorly on the validation set, there may be an overfitting problem, and further regularization or model adjustment is required.

[0101] Its technical effect is: by inputting the validation set into the trained model and calculating indicators such as accuracy, precision, and recall, a comprehensive evaluation of the model performance is achieved, which can effectively avoid the overfitting and underfitting problems of the model, and provide feedback for adjusting hyperparameters or modifying the network architecture. It not only improves the accuracy, stability, and robustness of the model, but also provides an important basis for the continuous optimization of the model.

[0102] In a preferred embodiment of the present invention, in the above-mentioned deep learning-based CT image analysis and tumor prediction method, the patient's CT image data is input into an optimized second-generation efficient convolutional neural network tumor classification model to obtain a classification result, and the existence, category or malignancy of the tumor is predicted according to the classification result, including: preprocessing the patient's CT image data; inputting the preprocessed CT image data into an optimized second-generation efficient convolutional neural network tumor classification model to perform shallow feature extraction, deep feature extraction, attention enhancement and classification prediction to obtain a prediction result representing the probability distribution of the model for each category. Select the category with the highest probability As the final prediction, if , then it is judged that the tumor exists; according to category c pred The specific meaning of the output prediction result is used, and the usage probability p is used cpred as the confidence level of the prediction.

[0103] After obtaining the prediction results, the classification results can be visualized, for example, high-risk areas can be marked through heat maps to assist doctors in locating the tumor; the classification results can be integrated into an easy-to-interpret report, including: whether the tumor exists, tumor type (benign / malignant), degree of malignancy (if applicable) and confidence score; if there is a deviation between the patient image prediction result and the actual diagnostic label, the result analysis is required to evaluate whether the model needs to be further optimized, for example, incremental training through more labeled data.

[0104] Its technical effect is: by inputting the patient's CT image data into the optimized second-generation efficient convolutional neural network tumor classification model, combining shallow and deep feature extraction, attention enhancement and probability prediction mechanism, it can accurately predict the existence, category and malignancy of the tumor. It not only provides accurate tumor classification results, but also provides confidence for doctors, improves the reliability and efficiency of diagnosis, and has strong clinical application value.

[0105] The second embodiment of the present invention provides a CT image analysis and tumor prediction system based on deep learning, which includes: a data acquisition module, which is used to collect CT image data and annotation information of tumor patients to obtain a medical image data set, and preprocess the medical image data set to obtain a high-quality CT image data set; a model building module, which is used to combine neural structure search technology and composite model expansion method to establish a second-generation efficient convolutional neural network tumor classification model; a model improvement module, which is used to add a coordinate attention module to the second-generation efficient convolutional neural network tumor classification model to obtain an improved second-generation efficient convolutional neural network tumor classification model; an optimization module, which is used to train and optimize on the basis of the improved second-generation efficient convolutional neural network tumor classification model; a result output module, which is used to input the patient's CT image data into the optimized second-generation efficient convolutional neural network tumor classification model to obtain a classification result, and predict the existence, category or malignancy of the tumor according to the classification result.

[0106] The computer program product of the deep learning-based CT image analysis and tumor prediction method and device provided in the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the method in the previous method embodiment. The specific implementation can be found in the method embodiment, which will not be repeated here.

[0107] Specifically, the storage medium can be a general storage medium, such as a mobile disk, a hard disk, etc. When the computer program on the storage medium is run, it can execute the above-mentioned deep learning-based CT image analysis and tumor prediction method, thereby significantly improving the analysis efficiency of CT images and the accuracy of tumor diagnosis.

[0108] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0109] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A CT image analysis and tumor prediction method based on deep learning, characterized in that: include: Collecting CT image data and annotation information of tumor patients to obtain a medical image data set, and preprocessing the medical image data set to obtain a high-quality CT image data set; Combining neural structure search technology and composite model expansion methods, a second-generation efficient convolutional neural network tumor classification model was established; A coordinate attention module and a hard smoothing activation function are added to the second-generation efficient convolutional neural network tumor classification model to obtain an improved second-generation efficient convolutional neural network tumor classification model; Based on the improved second-generation efficient convolutional neural network tumor classification model, training and optimization are performed; The patient's CT image data is input into the optimized second-generation efficient convolutional neural network tumor classification model to obtain the classification results, and the existence, category or malignancy of the tumor is predicted based on the classification results.

2. The method for CT image analysis and tumor prediction based on deep learning according to claim 1, characterized in that: The collecting of CT image data and annotation information of tumor patients to obtain a medical image data set, and preprocessing the medical image data set to obtain a high-quality CT image data set include: Collect CT image data of tumor patients with tumor types in designated locations and pathological diagnosis reports of tumor patients to obtain the specific location and pathological type of CT images and lesions; Annotate the CT image using a medical image annotation tool, generate a corresponding annotation mask, match the annotation information with the corresponding CT image, and obtain a medical image dataset; Using adaptive threshold segmentation and morphological operations to remove metal artifacts or motion artifacts in the CT image; Standardizing the different CT images to a fixed resolution using linear interpolation; Using a non-local means algorithm to perform smoothing and noise reduction on the CT image; Using a Laplacian operator to enhance the boundary of the lesion area in the CT image; The processed CT image data is converted into a format for a deep learning model and stored according to tumor type, stage or organ classification to obtain a high-quality CT image dataset.

3. The CT image analysis and tumor prediction method based on deep learning according to claim 1, characterized in that: The second generation of efficient convolutional neural network tumor classification model is established by combining neural structure search technology and composite model expansion method, including: Set the search space for the second-generation efficient convolutional neural network tumor classification model. The shallow layer uses a fused moving bottleneck convolution module to quickly extract low-level features from the input CT image data. The deep layer uses a deep separable moving bottleneck convolution module to extract features using depthwise separable convolution and attention mechanism. The fused moving bottleneck convolution module includes a common convolution kernel with a width of 3 pixels and a height of 3 pixels; The depth separation moving bottleneck convolution module includes a dimensionality increase convolution kernel with a width of 1 pixel and a height of 1 pixel, a depth separable convolution kernel with a width of 3 pixels and a height of 3 pixels, and a dimensionality reduction convolution kernel with a width of 1 pixel and a height of 1 pixel; Taking classification accuracy and computational cost as optimization goals, an objective function maximize Accuracy(Net(d,w,r))-λ·Cost(Net) is set for automatically finding the optimal network structure, where d is depth, w is width, r is input resolution, Cost(Net) is resource cost, and λ is a weight for balancing classification performance and computational cost, to obtain the optimal configuration of the second-generation efficient convolutional neural network tumor classification model; The composite scaling factor φ is set to make the second generation efficient convolutional neural network tumor classification model proportionally expand d=α in depth d, width w and input resolution r. φ , w=β φ , r = γ φ , establish the constraint relationship α·β of the parameters of the second generation efficient convolutional neural network tumor classification model 2 γ 2 ≈2,α≥1,β≥1,γ≥1, the depth d, width w and input resolution r of the neural network are uniformly scaled through the composite model expansion method to optimize the second-generation efficient convolutional neural network tumor classification model; The structure of the second-generation efficient convolutional neural network tumor classification model after compound scaling adjustment is: in, Indicates module stacking, FL i is the calculation function of the i-th convolution module, X is the input feature map, H i is the height of the input image, W i is the width of the input image, C i is the number of channels in each layer.

4. The method for CT image analysis and tumor prediction based on deep learning according to claim 3, characterized in that: The improved second-generation efficient convolutional neural network tumor classification model obtained by adding a coordinate attention module and a hard smoothing activation function to the second-generation efficient convolutional neural network tumor classification model includes: Adding a hard smoothing activation function to the fused mobile bottleneck convolution module; Adding an attention module and a hard smoothing activation function to the depth separation moving bottleneck convolution module; The attention module globally pools the feature map input to the depth separation moving bottleneck convolution module, extracts features, concatenates and maps them to a low-dimensional space, and then returns to the original space, and enhances the spatial dimension of the input feature map in a weighted manner; The hard smoothing activation function translates, truncates and normalizes the feature map input into the fused moving bottleneck convolution module or the deep separation moving bottleneck convolution module, and outputs the complex texture and morphological features of the tumor.

5. The method for CT image analysis and tumor prediction based on deep learning according to claim 4, characterized in that: The adding of a hard smoothing activation function to the fused mobile bottleneck convolution module comprises: The shallow input feature map of the fused mobile bottleneck convolution module is X1; The common convolution kernel is combined with the hard smoothing activation function to obtain the fused output Z of the fused mobile bottleneck convolution module. low =Hard-Swish(W1*X1+b1), where W1 is the weight of the shallow input feature map and b1 is the bias term; The formula of the hard smooth activation function is x is the element value of the shallow input feature map, ReLU6(x) is the activation function, and the input is truncated within the interval [0,6]. The formula is ReLU6(x)=min(max(0,x),6), which means that the element value of the shallow input feature map is shifted so that some negative values ​​participate in the activation.

6. The method for CT image analysis and tumor prediction based on deep learning according to claim 4, characterized in that: The adding of the attention module to the depth separation mobile bottleneck convolution module comprises: The depth input feature map of the deep separation moving bottleneck convolution module is For the feature map Perform global pooling in the vertical and horizontal directions respectively to obtain the horizontal encoding features and vertical encoding features Among them, c is the channel index, w is the horizontal position of the feature map, and h is the vertical position of the feature map; The horizontal encoding feature Y w and the vertical encoding feature Y h Through the cascade operation fusion, the compact feature representation Z = ReLU (W1 · [Y h ; Y w ]+b1), where the compact feature representation W1 is the dimension reduction weight, b1 is the bias parameter, and ReLU(·) is the nonlinear activation function; Map the compact feature representation Z back to the horizontal direction to obtain the horizontal mapping feature Map back to the vertical direction to obtain the vertical mapping feature Use attention weights to weight features X CA (h,w,c)=X in (h,w,c)·g h (h,c)·g w (w,c), get the output features of the attention module 7. The CT image analysis and tumor prediction method based on deep learning according to claim 1, characterized in that: The training and optimization based on the improved second-generation efficient convolutional neural network tumor classification model includes: Dividing the high-quality CT image dataset into a training dataset and a validation set; The training data set is input into the improved second-generation efficient convolutional neural network tumor classification model for forward propagation to obtain a prediction output Calculate the cross entropy loss function Where y is the predicted output The corresponding label, C is the number of categories, y i is the i-th label of the true label, is the predicted probability for the i-th category; Perform back propagation and update the model parameters by calculating the gradient of the loss function in, is the updated parameter, η1 is the learning rate, is the cross entropy loss function About weight W l The gradient of Adjust weights through the Adam optimization algorithm Among them, W t is the weight after the tth iteration, m t is the momentum of the gradient, v t is the average value of the square of the gradient, ò is a constant, and η2 is the learning rate; The validation set is used to validate the trained second-generation efficient convolutional neural network tumor classification model.

8. The method for CT image analysis and tumor prediction based on deep learning according to claim 7, characterized in that: The using the validation set to validate the trained second-generation efficient convolutional neural network tumor classification model includes: Loading the final parameters of the second-generation efficient convolutional neural network tumor classification model after training; The validation set is input into the improved second-generation efficient convolutional neural network tumor classification model for forward propagation to obtain a prediction output Output the prediction Compare with the true label of the sample, calculate the accuracy, precision, and recall rate. If the performance of the validation set does not meet expectations, it is determined that the hyperparameters of the model need to be adjusted or the network architecture needs to be modified, and the parameters need to be adjusted to obtain the model.

9. The CT image analysis and tumor prediction method based on deep learning according to claim 1, characterized in that: The patient's CT image data is input into the optimized second-generation efficient convolutional neural network tumor classification model to obtain a classification result, and the existence, category or malignancy of the tumor is predicted according to the classification result, including: Preprocessing the patient's CT image data; The pre-processed CT image data is input into the optimized second-generation efficient convolutional neural network tumor classification model to perform shallow feature extraction, deep feature extraction, attention enhancement and classification prediction to obtain the prediction results representing the probability distribution of each category by the model. Select the category with the highest probability As the final prediction, if , then the tumor is judged to exist; According to category c pred The specific meaning of the output prediction results and use the probability as the confidence level of the prediction.

10. A CT image analysis and tumor prediction system based on deep learning, characterized in that: include: A data acquisition module is used to collect CT image data and annotation information of tumor patients to obtain a medical image data set, and pre-process the medical image data set to obtain a high-quality CT image data set; The model building module is used to combine neural structure search technology and composite model expansion methods to build a second-generation efficient convolutional neural network tumor classification model; A model improvement module, used for adding a coordinate attention module to the second-generation efficient convolutional neural network tumor classification model to obtain an improved second-generation efficient convolutional neural network tumor classification model; An optimization module, used for training and optimizing based on the improved second-generation efficient convolutional neural network tumor classification model; The result output module is used to input the patient's CT image data into the optimized second-generation efficient convolutional neural network tumor classification model to obtain the classification results, and predict the existence, category or malignancy of the tumor based on the classification results.

Images