Deep learning-based CT image feature extraction and analysis method

Abstract

The invention relates to the field of CT (Computed Tomography) images, and discloses a deep learning-based CT image feature extraction and analysis method, which comprises the following steps of: performing de-noising processing, image normalization and data enhancement on an input head CT image; using a convolutional neural network to carry out preliminary detection of a focus area on the preprocessed CT image; using a generative adversarial network to carry out fine segmentation on the preliminarily detected lesion area; carrying out feature extraction on the lesion region after fine segmentation, and inputting the extracted features into a classifier to carry out lesion type classification; post-processing the classification result, and superposing the segmented focus area on the original CT image to generate an image with a focus mark; and evaluating the model, carrying out quantitative analysis on the model by adopting cross validation, and carrying out iterative optimization according to an evaluation result. The method has the advantages that stable extraction of the focus area is ensured, and the diagnosis efficiency and reliability are improved.

Description

Technical Field

[0001] The present invention relates to the field of CT images, and specifically to a method for CT image feature extraction and analysis based on deep learning. Background Art

[0002] Computed tomography (CT), as an important means of medical imaging, plays a key role in the diagnosis of head diseases. In the process of automatic analysis of head CT images, accurate segmentation of the lesion area is a key step affecting the diagnostic accuracy. However, in clinical applications, the gray-scale distributions of common intracranial hemorrhage, brain tumors and other lesions in CT images are complex, the boundaries are blurred, and they are easily affected by artifacts, noise and individual anatomical differences, resulting in difficulty for traditional segmentation methods to accurately extract the lesion area. Especially in the emergency environment, doctors need to quickly identify the lesions within a short time, while the existing deep learning-based segmentation methods often rely on large-scale, high-quality labeled data sets, and the trained models have insufficient generalization ability when facing a small amount of data or low-quality images, affecting clinical applicability. Therefore, aiming at the problem of the decline in the robustness of the model due to uneven data quality during the lesion segmentation of emergency head CT images, it is necessary to design an enhanced feature extraction and analysis method based on deep learning to improve the diagnostic efficiency and reliability. Summary of the Invention

[0003] Aiming at the deficiencies of the prior art, the present invention provides a method for CT image feature extraction and analysis based on deep learning, which has the advantages of ensuring the stable extraction of the lesion area and improving the diagnostic efficiency and reliability, and solves the problems in the above background art.

[0004] To achieve the above purpose of ensuring the stable extraction of the lesion area and improving the diagnostic efficiency and reliability, the present invention provides the following technical solution: A method for CT image feature extraction and analysis based on deep learning, including the following steps:

[0005] S1: Denoise the input head CT image, perform image normalization and data augmentation.

[0006] Preferably, the S1 further includes using Gaussian filtering, median filtering, bilateral filtering, deep learning autoencoders and non-local similarity methods to remove the noise and artifacts in the CT image while maintaining the detail and edge information, and performing image rotation, translation, scaling, flipping, brightness adjustment and noise simulation through pixel value scaling, normalization processing, contrast adjustment and adaptive histogram equalization to simulate different exposure conditions, electronic noise and discrete noise.

[0007] S2: Use a convolutional neural network to perform a preliminary detection of the lesion area on the preprocessed CT image.

[0008] Preferably, the step S2 further includes using a deep CNN network to extract key features of the lesion area by adjusting the parameters of the convolutional layer, pooling layer, and fully connected layer, adjusting the CT image to a fixed size by cropping or scaling, using multi-channel input to enhance the expression ability of lesion features, using multi-layer convolutional kernels to extract texture, edges, and shapes, predicting the possibility of the lesion area through the Logistic function, generating lesion candidate areas in combination with the region proposal network, and using an object detection algorithm to frame and output a probability score to generate a probability map.

[0009] Preferably, the step of using multi-layer convolutional kernels to extract texture, edges, and shapes and predicting the possibility of the lesion area through the Logistic function includes using an evaluation model constructed by the Logistic regression analysis method to conduct an overall risk assessment of the possibility of predicting the lesion area:

[0010] The exponential equation of Logistic is:

[0011] In the formula, P is the probability of predicting the possibility of the lesion area, x1 is the texture, x2 is the edge, x3 is the shape, T1, T2, and T3 are the regression coefficients of each variable, and b is the constant term.

[0012] S3: Use a generative adversarial network to perform fine segmentation on the preliminarily detected lesion area.

[0013] Preferably, the step S3 further includes using a generative adversarial network, where the generator performs fine segmentation of the lesion area, the discriminator evaluates the authenticity of the segmentation result, uses the preliminarily detected lesion area as the GAN input, uses real lesion area annotation data as the supervision signal, and uses a multi-scale discriminator to evaluate the global and local segmentation effects simultaneously.

[0014] S4: Extract features from the finely segmented lesion area and input the extracted features into a classifier for classifying the lesion type.

[0015] Preferably, the step S4 further includes combining deep learning and traditional feature extraction techniques to classify the lesion area, using ResNet, DenseNet, and VGG to extract high-level features, combining GLCM, LBP, and HOG to extract statistical features, simultaneously enhancing multi-scale information with the help of a feature pyramid network or self-attention mechanism, optimizing the feature representation through PCA or t-SNE dimensionality reduction, and standardizing the data using Batch Normalization. In the classification stage, use ResNet, EfficientNet, combine with random forest, XGBoost, and use Focal Loss to address class imbalance.

[0016] S5: Post-process the classification results, and overlay the segmented lesion area on the original CT image to generate an image with lesion markings.

[0017] Preferably, S5 further includes using transparency control technology to semi-transparently overlay the lesion area on the original CT image, using Canny edge detection to extract the lesion contour, and highlighting it on the original CT image, saving the processed image in DICOM format, and generating multi-view images and an automatic diagnosis report.

[0018] S6: Evaluate the model, perform quantitative analysis on the model using cross-validation, and perform iterative optimization according to the evaluation results.

[0019] Preferably, S6 further includes using K-fold cross-validation or leave-one-out method to evaluate the stability of the model, using an independent test data set not involved in training to evaluate the generalization ability of the model, statistically analyzing the prediction confidence distribution, checking low-confidence samples, analyzing misclassified or mis-segmented samples, finding common error types, using network pruning or knowledge distillation to reduce the model complexity, using Bayesian optimization to adjust hyperparameters, performing transfer learning on data of specific lesion types, training the model using adversarial samples, using a decentralized training method, and integrating data from multiple hospitals.

[0020] Compared with the prior art, the present invention provides a method for CT image feature extraction and analysis based on deep learning, having the following beneficial effects:

[0021] Through a multi-stage deep learning model, including using a convolutional neural network for preliminary lesion detection, a generative adversarial network for fine segmentation, and the combination of a feature extraction and classification model, the present invention improves the recognition ability for complex lesion areas. Data augmentation and denoising processing are used to enhance the adaptability of the model to low-quality and small-sample data and improve the generalization ability. The cross-validation and iterative optimization mechanisms are used to enable the model to maintain stable performance under different individual anatomical differences and noise interferences. Aiming at the problem of the decline in model robustness due to uneven data quality during the lesion segmentation process of emergency head CT images, this method can quickly and accurately mark the lesion area in CT images, improve the diagnosis efficiency, and provide reliable auxiliary decision-making support for emergency and clinical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic diagram of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0024] The present invention provides a technical solution: a method for CT image feature extraction and analysis based on deep learning, including the following steps:

[0025] S1: Perform denoising processing, image normalization, and data augmentation on the input head CT image.

[0026] Adopt methods such as Gaussian filtering, median filtering, or bilateral filtering to remove noise and artifacts in the CT image. Use the Gaussian function for smoothing to reduce the influence of high-frequency noise. Replace the central pixel value with the median value of the pixels within the window to remove salt-and-pepper noise. At the same time, consider the similarity of pixel positions and gray values to achieve edge-protective denoising. Based on a deep learning autoencoder, train the model to learn the features of the noise-free image and remove the noise part. Based on non-local similarity, perform weighted averaging on similar pixel blocks to remove random noise and preserve details. Scale the pixel values of the CT image to the interval [0, 1] or [-1, 1] to reduce the numerical span of the data and improve the convergence speed of the model. Normalize the pixel values of the image with a mean of 0 and a standard deviation of 1 to eliminate the influence of different devices and scanning conditions. Adjust the gray distribution of the image to enhance the contrast and improve the distinguishability of the lesion area. Perform histogram equalization within a local area to prevent detail loss caused by over-enhancement. Randomly rotate the image within a certain range (such as -30 to 30) to simulate different scanning angles. Randomly translate pixels in the horizontal or vertical direction to enhance the model's adaptability to different displacements. Randomly zoom in or out the image to enhance the model's ability to recognize lesions of different sizes. Perform horizontal or vertical flipping to expand the diversity of data samples. Increase or decrease the brightness of the image to simulate different exposure conditions. Adjust the contrast of the image to make the distinction between the lesion and normal tissue more obvious. Simulate the electronic noise that may exist in real CT images to improve the anti-interference ability of the model. Simulate the discrete noise during the CT scanning process to improve the robustness of the model.

[0027] Filtering algorithms can effectively remove high-frequency noise and artifacts in CT images, making the lesion areas clearer. Methods such as bilateral filtering and autoencoder denoising can remove noise while preserving edge information, preventing the blurring of lesion boundaries. Reducing the impact of noise on the feature extraction of deep learning models improves the accuracy of lesion recognition. Normalization reduces data bias caused by different CT scanning devices and imaging parameters, improving the generalization ability of the model on different datasets. The numerical range of the normalized image data is stable, which helps the stability of gradient descent, accelerates the convergence of the model, and improves the training efficiency. Histogram equalization can enhance the lesion areas in low-contrast images, improving the doctor's interpretation ability and the model's detection ability. Geometric transformation and color enhancement increase the diversity of training data, enabling the model to adapt to different CT scanning conditions and lesion morphologies. Data augmentation expands the training set, preventing the model from overfitting to specific samples and improving its performance on unseen data. Through noise injection, the model can learn to maintain stable feature extraction capabilities at different noise levels, improving its robustness in actual clinical applications.

[0028] S2: Use a convolutional neural network to perform a preliminary detection of the lesion area on the preprocessed CT image.

[0029] Adopt a classic convolutional neural network (CNN) architecture for the preliminary detection of the lesion area. Design a deep network adapted to CT images, and adjust the parameters of the convolutional layer, pooling layer, and fully connected layer so that it can extract the key features of the lesion area. Adopt a fixed-size input, such as cropping or scaling the CT image to a standard size of 224×224 or 512×512 to adapt to the CNN network structure. Adopt a multi-channel input, such as inputting CT images at different levels into different channels to enhance the ability to express lesion features. Adopt multi-layer convolutional kernels (such as 3×3 or 5×5) to extract the texture, edge, and shape information of the lesion area. Use batch normalization to improve the training stability and accelerate the convergence speed. Adopt a non-linear activation function to enhance the expression ability of feature extraction. Use max pooling or average pooling to reduce the size of the feature map, reduce the computational amount, and retain key information at the same time. Enhance the ability to express the features of the lesion area through the attention mechanism and reduce background interference. Use a fully connected layer or a 1×1 convolutional layer for feature mapping, and predict the possibility of the lesion area through the Logistic function. Adopt a region proposal network to generate possible lesion candidate regions to improve the accuracy of the preliminary detection. Combine with an object detection algorithm to frame the lesion area and output a probability score. Generate a bounding box of the lesion area for annotating possible lesion regions. Generate a probability map representing the confidence distribution of the lesion area for subsequent fine segmentation.

[0030] Automatically extract the features of the lesion area through CNN to improve the detection speed and reduce the workload of manual annotation by doctors. Adopt methods such as feature pyramids to improve the model's detection ability for lesions of different sizes. Extract lesion features through a multi-layer convolutional network to improve the robustness of detection, reduce false positives and missed detections. Adopt an attention mechanism to make the model pay more attention to the lesion area, reduce background interference, and improve detection accuracy. Through methods such as input image normalization and batch normalization, reduce the data distribution differences caused by different CT scanning devices and improve the model's generalization ability. Adopt multi-channel input, combine information of different window widths / window levels, and improve the detection ability for complex lesions. Generate lesion candidate regions to provide a preliminary target area for subsequent fine segmentation and reduce computational redundancy. Through probability maps or heatmaps, guide the classifier to perform more accurate lesion type recognition and improve the overall diagnostic accuracy.

[0031] S3: Use a generative adversarial network to perform fine segmentation on the preliminarily detected lesion area.

[0032] Adopt a generator-discriminator architecture, where the generator is responsible for generating the finely segmented lesion area, and the discriminator is used to judge the authenticity of the generated result. The generator uses U-Net, DeepLabV3+ or ResNet as the basic structure for pixel-level segmentation of the lesion area. The discriminator uses PatchGAN or a global / local discriminant network to evaluate whether the segmentation result matches the real lesion area. Use the preliminarily detected lesion area as the input of the GAN to reduce background interference and improve the segmentation accuracy. Use real lesion area annotation data as the supervision signal to guide the GAN to learn high-precision segmentation. Adopt skip connections to enhance the segmentation edge details and prevent information loss. Adopt residual modules to increase the depth of the model and enhance the feature expression ability of the lesion area. Use upsampling or transposed convolution to restore the high-resolution information of the segmented area. Adopt a multi-scale discriminator to evaluate the global and local segmentation effects simultaneously to ensure the structural rationality of the lesion area. Adopt PatchGAN to improve the segmentation accuracy of the boundary area by discriminating small regions.

[0033] The loss functions include:

[0034] Adversarial loss: The generator minimizes the adversarial loss to make the segmentation result closer to the real lesion area.

[0035] Cross-entropy loss or Dice loss: Measure the overlap degree between the segmented area and the real annotation to improve the segmentation accuracy.

[0036] Total variation loss: Constrain the smoothness of the generated segmented area and reduce the influence of noise.

[0037] Adopt an alternating training strategy, that is, first train the discriminator and then train the generator to ensure the stable training of the GAN. Use data augmentation (such as flipping, rotation, scaling) to expand the training set and improve the generalization ability of the GAN. Adopt semi-supervised or weakly-supervised learning to make full use of unlabeled data and improve the robustness of the segmentation model.

[0038] Through adversarial learning, the generator can learn pixel-level segmentation closer to the real lesion area and improve the boundary clarity. Adopt skip connections and residual modules to make the details of the segmented area better preserved and reduce the boundary blurring phenomenon. Through the feedback of the discriminator, the generator can continuously optimize the segmentation results, reduce false positives and false negatives, and improve the stability of segmentation. Adopt multi-scale discriminators and PatchGAN to improve the segmentation ability of small lesions and reduce the interference of the background area. Adopt total variation loss to make the morphology of the lesion area more in line with medical characteristics and reduce the generation of non-real structures. Through semi-supervised learning, combined with unlabeled data, improve the adaptability of the model to different lesion morphologies. Use data augmentation and weakly-supervised strategies to improve the robustness of the GAN on different devices and different patient data. Through the alternating training strategy, ensure the training stability of the GAN and avoid mode collapse.

[0039] S4: Extract features from the finely segmented lesion area and input the extracted features into a classifier for lesion type classification.

[0040] Use deep networks such as ResNet, DenseNet, VGG to extract high-level features of the lesion area, including edge information, texture features, shape features, etc. Combine methods such as gray-level co-occurrence matrix (GLCM), local binary pattern (LBP), histogram of oriented gradients (HOG) to extract statistical features of the lesion area. Use feature pyramid networks or self-attention mechanisms to extract lesion information at different scales and improve the accuracy of classification. Adopt principal component analysis (PCA) or t-SNE to reduce the dimension of the extracted features, remove redundant information, and reduce the computational complexity. Use Batch Normalization for feature standardization to improve the separability of features. Use networks such as ResNet, EfficientNet as classifiers to classify the lesions as benign / malignant or multi-class. Establish a hyperplane in the high-dimensional feature space to achieve lesion type classification. Use a fully-connected neural network to classify and predict the feature vectors. Use methods such as random forest (RF), extreme gradient boosting (XGBoost) to improve the stability of classification. Use data augmentation to expand the lesion samples and enhance the generalization ability of the model to small-sample data. Use Focal Loss to solve the class imbalance problem and make the classifier pay more attention to difficult-to-classify samples. Adopt model fusion, such as integrating multiple CNN classifiers, and improve the classification accuracy through a voting mechanism.

[0041] Adopt deep neural networks and multi-scale feature extraction methods to improve the model's ability to express lesion features. Combine traditional image features and deep learning features to enhance the separability of lesions and improve the robustness of classification. Reduce redundant information through PCA dimensionality reduction and Batch Normalization to enhance classification stability. Employ an attention mechanism to enable the model to focus on the key features of the lesion area, reduce background interference, and improve classification accuracy. Expand training data through data augmentation to improve the model's generalization ability for lesions of different patients. Use Focal Loss to balance class imbalance and improve the model's recognition ability for rare lesion types. Adopt a dimensionality reduction method to reduce computational complexity and improve the model's inference speed. Use a lightweight model to enhance the model's deployment ability on edge devices.

[0042] S5: Post-process the classification results and overlay the segmented lesion area on the original CT image to generate an image with lesion markings.

[0043] Set a confidence threshold (such as 0.5 or 0.7) to screen out low-confidence classification results and reduce misjudgment. For multi-class classification tasks, use Sigmoid normalization to make the prediction results conform to the probability distribution and improve the stability of class determination. Use Platt Scaling or temperature scaling to calibrate the probabilities output by the model to reduce overfitting of the classification model. Use morphological operations such as dilation, erosion, opening, and closing to optimize the edges of the lesion area to make it smoother and more coherent. Use connected component labeling to remove small or discrete pseudo-lesion areas and improve the accuracy of segmentation. For over-segmented lesion areas, use region growing or conditional random fields for merging to improve integrity. Visualize the lesion area with pseudo-colors (such as red, blue, heat maps, etc.) to enhance the distinguishability of the lesions. Use transparency control technology to semi-transparently overlay the lesion area on the original CT image to retain the underlying anatomical structure information. Use Canny edge detection to extract the lesion contour and highlight it on the original CT image for doctors to quickly locate the lesion position. Save the processed image in DICOM format to be compatible with the hospital PACS and support doctors' image viewing. Generate axial, coronal, and sagittal lesion visualization images to help doctors analyze the lesion morphology from multiple angles. Automatically generate a report containing information such as lesion type, size, and location to assist doctors in diagnosis.

[0044] Reduce misjudgment and improve the stability of the classification model through confidence threshold screening and probability calibration. Optimize the lesion segmentation boundary and improve the segmentation quality through morphological processing and connected component analysis. Make the lesion area more intuitive and improve the doctor's film reading efficiency by using methods such as pseudo-color mapping, transparency fusion, and edge highlighting. Provide more complete lesion information and assist doctors in comprehensive diagnosis by using multi-view display. Store in DICOM format to ensure that the images can be seamlessly integrated into the hospital information system for convenient access by doctors. Generate a medical report containing the classification results, reduce the workload of manual recording by doctors, and improve the automation of the diagnosis process.

[0045] S6: Evaluate the model, perform quantitative analysis on the model using cross-validation, and perform iterative optimization according to the evaluation results.

[0046] Use K-fold cross-validation or leave-one-out method to evaluate the stability of the model. During cross-validation, keep the proportion of samples in different categories consistent to reduce the impact of class imbalance on the evaluation results. Use an independent test dataset not involved in training to evaluate the generalization ability of the model and avoid overfitting. Calculate metrics such as accuracy, precision, recall, F1 score, ROC-AUC curve, etc. to comprehensively evaluate the classification ability of the model. Use Dice coefficient, IoU, Hausdorff distance, etc. to measure the accuracy of model segmentation. Statistically analyze the confidence distribution of predictions, check low-confidence samples, and improve decision reliability. Analyze misclassified or incorrectly segmented samples to find common error types. For the problem of class imbalance in the dataset, use Focal Loss, weighted loss function, or data augmentation techniques to optimize the model. Use network pruning or knowledge distillation to reduce the model complexity and improve the inference efficiency. Use Bayesian optimization or grid search to adjust hyperparameters. Perform transfer learning on data of specific lesion types to improve the model adaptability. Train the model with adversarial samples to improve the robustness of the model to input perturbations. Adopt a decentralized training method to integrate data from multiple hospitals, enhance the model generalization ability, and avoid privacy leakage.

[0047] Through cross-validation and external test set validation, ensure the stability of the model on different data sets and reduce overfitting. Adopt error sample analysis and class imbalance handling to improve the recognition ability of rare lesion types. Use a variety of performance metrics for quantitative analysis to comprehensively evaluate the model effect and ensure high precision in classification and segmentation tasks. Optimize the lesion boundary segmentation using the Hausdorff distance to improve the accuracy of doctors' judgment on lesion morphology. Adopt techniques such as model pruning and knowledge distillation to reduce the computational cost and make the model easier to deploy in a clinical environment or on mobile devices. Through hyperparameter optimization, improve the training efficiency, reduce the experimental cost, and speed up the model tuning process. Adopt adversarial training to enhance the model's adaptability to abnormal inputs and improve the reliability of clinical applications. Through the way of federated learning, share the model capabilities among multiple medical institutions to improve the adaptability and privacy protection level of the overall medical AI.

[0048] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0049] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A CT image feature extraction and analysis method based on deep learning, characterized in that, It includes the following steps: S1: Denoise, normalize, and augment the input head CT image; S2: Use a convolutional neural network to perform a preliminary detection of the lesion area on the preprocessed CT image; S3: Use a generative adversarial network to perform a fine segmentation of the preliminarily detected lesion area; S4: Extract features from the finely segmented lesion area and input the extracted features into a classifier for lesion type classification; S5: Post-process the classification result and superimpose the segmented lesion area on the original CT image to generate an image with lesion markings; S6: Evaluate the model, perform quantitative analysis on the model using cross-validation, and perform iterative optimization according to the evaluation results.

2. The CT image feature extraction and analysis method based on deep learning according to claim 1, characterized in that The S1 further includes using Gaussian filtering, median filtering, bilateral filtering, deep learning autoencoders, and non-local similarity methods to remove noise and artifacts in the CT image while preserving details and edge information, and performing image rotation, translation, scaling, flipping, brightness adjustment, and noise simulation through pixel value scaling, normalization, contrast adjustment, and adaptive histogram equalization to simulate different exposure conditions, electronic noise, and discrete noise.

3. The CT image feature extraction and analysis method based on deep learning according to claim 1, characterized in that, The S2 further includes using a deep CNN network to extract key features of the lesion area by adjusting the parameters of the convolutional layer, pooling layer, and fully connected layer, adjusting the CT image to a fixed size by cropping or scaling, using multi-channel input to enhance the lesion feature expression ability, using multi-layer convolutional kernels to extract texture, edges, and shapes, predicting the possibility of the lesion area through the Logistic function, generating lesion candidate areas in combination with the region proposal network, and using an object detection algorithm for bounding and outputting probability scores to generate a probability map.

4. The CT image feature extraction and analysis method based on deep learning according to claim 3, characterized in that, The step of using multi-layer convolutional kernels to extract texture, edges, and shapes and predicting the possibility of the lesion area through the Logistic function includes using an evaluation model constructed by the Logistic regression analysis method to perform an overall risk assessment of the possibility of predicting the lesion area: The exponential equation of Logistic is as follows: where P is the probability of the possibility of predicting the lesion area, x1 is the texture, x2 is the edge, x3 is the shape, T1, T2, and T3 are the regression coefficients of each variable, and b is the constant term.

5. The CT image feature extraction and analysis method based on deep learning according to claim 1, characterized in that, The S3 further includes using a generative adversarial network, where the generator performs a fine segmentation of the lesion area, the discriminator evaluates the authenticity of the segmentation result, uses the preliminarily detected lesion area as the GAN input, uses real lesion area annotation data as the supervision signal, and uses a multi-scale discriminator to evaluate the global and local segmentation effects simultaneously.

6. The CT image feature extraction and analysis method based on deep learning according to claim 1, characterized in that S4 further includes combining deep learning and traditional feature extraction techniques to classify the lesion area, extracting high-level features using ResNet, DenseNet, and VGG, and combining GLCM, LBP, and HOG to extract statistical features. At the same time, it enhances multi-scale information with the help of a feature pyramid network or self-attention mechanism, optimizes the feature representation through PCA or t-SNE dimensionality reduction, and standardizes the data using Batch Normalization. In the classification stage, ResNet and EfficientNet are used, combined with random forest and XGBoost, and Focal Loss is used to address class imbalance.

7. The CT image feature extraction and analysis method based on deep learning according to claim 1, wherein S5 further includes adopting transparency control technology to semi-transparently overlay the lesion area on the original CT image, using Canny edge detection to extract the lesion contour and highlighting it on the original CT image, saving the processed image in DICOM format, and generating multi-view images and an automatic diagnosis report.

8. The CT image feature extraction and analysis method based on deep learning according to claim 1, characterized in that, S6 further includes using K-fold cross-validation or leave-one-out method to evaluate the stability of the model, using an independent test dataset not involved in training to evaluate the generalization ability of the model, statistically analyzing the prediction confidence distribution, checking low-confidence samples, analyzing misclassified or missegmented samples, finding common error types, reducing the model complexity using network pruning or knowledge distillation, adjusting hyperparameters using Bayesian optimization, performing transfer learning on data of specific lesion types, training the model with adversarial samples, adopting a decentralized training method, and integrating data from multiple hospitals.

Images