Medical image computer-aided analysis method based on deep learning

Abstract

The invention relates to the field of artificial intelligence, in particular to a medical image computer-aided analysis method based on deep learning, and aims to solve the problems that an existing medical image analysis method is low in high-resolution image processing efficiency, insufficient in tiny focus recognition precision, weak in model generalization ability and insufficient in multi-modal image fusion. According to the method, a lightweight multi-scale feature extraction network is constructed to improve the high-resolution image processing efficiency, a fine-grained lesion recognition module is introduced to improve the detection precision of a tiny lesion, and a self-adaptive regularization strategy is adopted to enhance the model generalization ability. And a multi-modal deep fusion mechanism is designed to make full use of complementary information of different modal images. According to the invention, medical image analysis which is more efficient, more accurate, higher in generalization ability and capable of effectively fusing multi-modal information can be realized, so that clinical application of deep learning in the field of medical images is promoted.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of artificial intelligence, and specifically relates to a medical image computer-aided analysis method based on deep learning. BACKGROUND

[0002] Medical image analysis plays an indispensable role in disease diagnosis, treatment planning, and prognosis evaluation. With the development of modern medical technology, various imaging modalities such as CT, MRI, and ultrasound have become increasingly popular, generating a vast amount of medical image data. Efficient and accurate analysis of these complex and information-rich image data is crucial to improving the quality and efficiency of medical services. In this context, computer-aided analysis methods have been introduced into the field of medical imaging to assist doctors in making more objective and accurate diagnoses.

[0003] Among them, the medical image computer-aided analysis method based on deep learning can automatically learn and extract high-dimensional features from raw image data through the construction of a multi-layer neural network, thereby realizing the detection, segmentation, classification, and quantitative analysis of lesions. This method effectively avoids the tediousness and subjectivity of traditional manual feature extraction, significantly improves the automation level and objectivity of medical image analysis, and has great potential in the auxiliary diagnosis of various diseases.

[0004] However, in the prior art, although deep learning has made significant progress in the field of medical image analysis, it still faces many challenges. Traditional deep learning models often have high computational complexity, large memory consumption, and long training periods when dealing with high-resolution or three-dimensional medical images, making it difficult to achieve efficient real-time processing. For small lesions with various shapes and blurred boundaries, the recognition accuracy and recall rate of the model still need to be improved. At the same time, under limited labeled data, the generalization ability of existing models is often limited, and overfitting phenomenon often occurs, leading to performance degradation when facing data from different devices and different patient groups. In addition, existing methods lack effective cross-modal feature alignment and deep fusion mechanisms when integrating multi-modal medical image information, making it difficult to fully utilize the complementary information between different image modalities, thereby affecting the final analysis accuracy and robustness. These problems seriously restrict the further clinical application and promotion of deep learning in the field of medical image computer-aided analysis, therefore, there is an urgent need for a medical image computer-aided analysis method based on deep learning that can overcome the above-mentioned defects of existing technology, provide more efficient, more accurate, and more generalizable analysis, and effectively fuse multi-modal information. SUMMARY

[0005] To solve the problems of high computational complexity, large memory consumption, long training period, difficulty in achieving efficient real-time processing, and the need to improve the recognition accuracy and recall rate of small lesions with different shapes and fuzzy boundaries, as well as the challenges of limited model generalization ability prone to overfitting under limited labeled data, and the lack of effective cross-modal feature alignment and deep fusion mechanism when fusing multi-modal medical image information, the present application proposes a deep learning-based medical image computer-aided analysis method.

[0006] The embodiments of the present application provide a deep learning-based medical image computer-aided analysis method, which constructs a lightweight multi-scale feature extraction network to improve the processing efficiency of high-resolution images, and introduces a fine-grained lesion recognition module to improve the detection accuracy of small lesions. The method further uses an adaptive regularization strategy to enhance the model generalization ability, and designs a multi-modal deep fusion mechanism to fully utilize the complementary information of different modal images. In this way, the limitations of existing technologies in computational efficiency, recognition accuracy, model generalization ability, and multi-modal fusion can be effectively overcome, providing a more efficient, more accurate, and more generalizable medical image analysis solution that can effectively fuse multi-modal information, thereby promoting the clinical application and promotion of deep learning in the field of medical images.

[0007] According to one aspect of the present application, a deep learning-based medical image computer-aided analysis method is provided, which includes the following steps: Receiving medical image data to be analyzed, which can include single or multi-modal images; Pretreating and feature extracting the medical image data to obtain multi-scale image feature representation; Performing lesion recognition or segmentation based on the multi-scale image feature representation and a pre-trained deep learning model to generate preliminary analysis results; Post-processing and result evaluation of the preliminary analysis results to output the final computer-aided analysis report.

[0008] Compared with the prior art, the medical image computer-aided analysis method based on deep learning provided by the application can improve the high-resolution image processing efficiency by constructing a lightweight multi-scale feature extraction network, and can improve the detection accuracy of small lesions by introducing a fine-grained lesion recognition module. The method further adopts an adaptive regularization strategy to enhance the model generalization ability, and designs a multi-modal deep fusion mechanism to fully utilize the complementary information of different modal images. In this way, the limitations of the prior art in terms of computational efficiency, recognition accuracy, model generalization ability and multi-modal fusion can be effectively overcome, and a more efficient, more accurate, more generalizable and multi-modal information fusion medical image analysis scheme can be provided, thereby promoting the clinical application and promotion of deep learning in the medical image field.

[0009] The embodiments of the application will be described in detail below.

[0010] The step of receiving medical image data to be analyzed can include single or multi-modal image data, specifically including obtaining original medical images from medical image devices or data storage systems. The medical image data can cover multiple imaging modalities, such as computed tomography images, magnetic resonance imaging images, positron emission tomography images, ultrasound images, digital subtraction angiography images, and pathological section images, etc. Each modality of image data has its own information characteristics and clinical application focus. For example, computed tomography images are good at showing bone structure and calcified lesions, and magnetic resonance imaging images have an advantage in soft tissue contrast. Positron emission tomography images can provide functional metabolic information, while ultrasound images have the characteristics of real-time and no radiation. During the receiving process, the system performs necessary metadata analysis to obtain image acquisition parameters, patient information, scanning protocols and other auxiliary information, which are crucial for subsequent preprocessing and analysis. For multi-modal images, the system also performs image sequence matching and association to ensure that image data of different modalities can correctly correspond to the same lesion area of the same patient, laying a foundation for subsequent deep fusion. The accuracy of this step directly affects the effectiveness and reliability of subsequent analysis.

[0011] The step of preprocessing and feature extraction of the medical image data to obtain multi-scale image feature representation specifically includes a series of data normalization and information enhancement operations. The preprocessing operation aims to eliminate noise and artifacts in the original data, and to unify the image processing, including but not limited to the following aspects: Image format conversion and data normalization: converting the received image data of different formats into a standard format that can be processed by the system. The pixel values of the image are normalized, for example, mapped to the range of zero to one, to eliminate the intensity difference caused by different acquisition devices or protocols, and to improve the training stability and convergence speed of the model.

[0012] Image registration: For multi-modal image data, perform precise registration operations to spatially align images of different modalities, ensuring that corresponding pixels of the same anatomical location can be accurately overlapped in different modal images. The registration can use intensity-based or feature-based algorithms combined with affine transformation, elastic transformation and other geometric correction methods to compensate for spatial deviations caused by patient position, respiratory motion or scanning field differences.

[0013] Noise suppression and artifact removal: Use filtering algorithms such as Gaussian filtering, median filtering or non-local mean filtering to effectively suppress random noise in the image. At the same time, identify and remove metal artifacts, motion artifacts and other artifacts caused by equipment, patient motion or reconstruction process to purify image quality and ensure the accuracy of subsequent feature extraction.

[0014] Region of interest cropping and enhancement: According to prior knowledge or automatic detection algorithm, crop the region of interest containing potential lesions to reduce the interference of irrelevant background information and reduce the computational burden. Perform local contrast enhancement on the region of interest, such as using histogram equalization, adaptive histogram equalization and other techniques to highlight lesion details.

[0015] After completing the preprocessing, the system will perform core feature extraction to construct a lightweight multi-scale feature extraction network to efficiently obtain deep semantic information of the image. The design of the lightweight multi-scale feature extraction network aims to balance computational efficiency and feature expression ability, especially suitable for high-resolution or three-dimensional medical images. The network can include the following structures: Encoder-decoder architecture: Use classic encoder-decoder structures such as U-shaped networks, fully convolutional networks, etc. as the basic skeleton. The encoder part gradually extracts high-level semantic features and reduces the size of the feature map through a series of convolutional layers, pooling layers or stride convolutional layers, while increasing the receptive field. The decoder part fuses the low-level spatial information and high-level semantic information extracted by the encoder through upsampling layers and skip connections, gradually restores the spatial resolution of the feature map, and finally generates a fine segmentation map or feature map.

[0016] Multi-scale feature fusion module: Embed multi-scale feature fusion mechanism in the encoder or decoder. This can include a spatial pyramid pooling module that captures different context information through different scale pooling operations; or use a feature pyramid network structure to build a feature pyramid with rich semantic information from the encoder feature maps at different levels, effectively processing lesions of different sizes. Through dilated convolution, the receptive field can be expanded without increasing the number of parameters and computational complexity, capturing more extensive context information, especially suitable for identifying small lesions with fuzzy boundaries and diverse morphology. Dilated convolution has multi-branch parallel processing capability, which can capture context information at different scales.

[0017] Lightweight convolution module: To reduce computational complexity and memory consumption, lightweight design such as depthwise separable convolution, group convolution or channel attention mechanism is adopted. Depthwise separable convolution decomposes standard convolution into two steps of depthwise convolution and pointwise convolution, significantly reducing the number of parameters and computation. Group convolution further improves efficiency by grouping input feature channels for convolution. Channel attention mechanism can adaptively adjust the weights of different channels according to the importance of input features, making the network pay more attention to features related to lesions.

[0018] Adaptive attention mechanism: Channel attention module or spatial attention module is introduced to enable the network to dynamically focus on the areas in the image that are crucial for diagnosis, such as potential lesion areas. Channel attention mechanism enhances or suppresses features in different channels by learning the weights of each channel, while spatial attention mechanism emphasizes features in important areas by learning the weights of spatial positions. This mechanism is particularly effective in improving the detection accuracy of fine-grained lesions.

[0019] Feature dimension compression and enhancement: Bottleneck structure, residual connection and other methods are used to compress and enhance feature dimensions at different levels to reduce the occupation of computing resources while maintaining the amount of information, and to alleviate the problem of gradient vanishing.

[0020] The step of performing lesion recognition or segmentation based on the multi-scale image feature representation and the pre-trained deep learning model to generate preliminary analysis results is the core diagnostic part of the method. This step uses the trained deep learning model to infer the extracted multi-scale image features to achieve accurate recognition or segmentation of lesions. Specific implementation schemes can include: Lesion detection and positioning: For tasks that need to identify whether a specific lesion exists in the image and determine its position, a target detection model is used. The model can be based on single-stage detectors such as YOLO series or SSD series, or based on two-stage detectors such as Faster R-CNN. These models use the feature maps output by the lightweight multi-scale feature extraction network to predict the class and bounding box coordinates of the lesion. By introducing a fine-grained lesion recognition module, the detection ability for small and irregular lesions can be enhanced. This module can use a multi-scale feature fusion strategy to combine high-level semantic information and low-level spatial information, or design a specific anchor mechanism to adapt to small size lesions.

[0021] Lesion region segmentation: For tasks that require precise delineation of lesion boundaries, semantic segmentation or instance segmentation models are employed. The models can be based on U-Net, Deeplab series, etc. These models utilize decoders to gradually restore spatial resolution and combine skip connections to pass detailed information captured by encoders to decoders, thereby generating pixel-level lesion masks. The fine-grained lesion identification module performs well in accurately distinguishing lesions from normal tissues, especially in cases where lesion boundaries are ambiguous or have low contrast with the background. For example, boundary-aware loss functions or attention gate mechanisms can be introduced to improve boundary segmentation accuracy.

[0022] Lesion classification and differentiation: For tasks that require classifying detected or segmented lesions, such as benign and malignant classification, classification models are employed. The models can be a standalone classification network or a backend classification head of the detection / segmentation model. By learning morphological, textural, and functional features of lesions through deep learning models, accurate classification of lesions is achieved.

[0023] Multi-modal deep fusion mechanism: When dealing with multi-modal medical images, the method designs an efficient multi-modal deep fusion mechanism to fully utilize the complementary information between different modalities. The fusion mechanism can be performed at different stages: Early fusion: In the feature extraction stage, the original images or their preliminary features of different modalities are spliced or element-level fused, and then input into a unified deep learning model for processing.

[0024] Mid-level fusion: At the intermediate level of the network, the features extracted from different modalities are interactively fused. For example, cross-modal attention mechanisms can be used to guide the enhancement of features from one modality to another, or common embedding spaces can be learned to capture shared representations of different modalities. This helps to capture more complex cross-modal associations, especially in fine-grained lesion identification.

[0025] Late fusion: Different modalities are analyzed independently, and then the recognition or segmentation results of each modality are fused for decision-making. For example, voting mechanisms, weighted averaging, or meta-learning models can be used to integrate the prediction results of different modalities.

[0026] Pre-trained deep learning models: The deep learning models used can be pre-trained on large-scale public datasets to learn general medical image features, and then fine-tuned on specific datasets for the task. Pre-training helps the model achieve better performance with limited labeled data.

[0027] Adaptive regularization strategies: To enhance the generalization ability of the model with limited labeled data, a series of adaptive regularization strategies are employed, including but not limited to the following techniques: Data Augmentation: Randomly flipping, rotating, scaling, cropping, brightness adjusting, elastic deformation, etc. are applied to the training data to increase the diversity of training samples and reduce the model's dependence on specific data patterns.

[0028] Batch Normalization: Adding a batch normalization layer after each network layer to accelerate network training and serve as a certain regularization, reducing internal covariate shift.

[0029] Dropout: Randomly discarding a portion of neuron outputs during training to force the network to learn more robust feature representations and avoid overfitting.

[0030] Weight Decay: Adding an L2 norm penalty term for model weights in the loss function to limit the size of model parameters and prevent the model from being too complex.

[0031] Label Smoothing: Softening hard labels in classification tasks, such as adjusting one-hot encoded label values from zero and one to zero point one and zero point nine, to reduce model overconfidence and improve generalization ability.

[0032] Contrastive Learning: Learning discriminative feature representations of images using unsupervised or self-supervised methods, which can effectively improve model performance even in the case of scarce labeled data.

[0033] The step of post-processing and result evaluation of the preliminary analysis results to output the final computer-aided analysis report aims to refine and clinically transform the model's output, ensuring the accuracy, reliability, and interpretability of the final results. Specific steps include: Post-processing operations: Connected Component Analysis: Perform connected component analysis on the segmentation results to remove small, isolated false positive regions or merge adjacent true lesion fragments.

[0034] Morphological Operations: Use morphological operations such as opening, closing, dilation, and erosion to smooth lesion boundaries, fill holes, or remove burrs, making the segmentation results more consistent with anatomical shapes in medicine.

[0035] False Positive Reduction: Combine contextual information or statistical methods to further reduce false positives in detection or segmentation. For example, by analyzing the location, size, shape, and other characteristics of lesions, filter out atypical false positives.

[0036] Quantitative Analysis and Evaluation: Lesion Measurement: Accurately measure the identified or segmented lesions, including volume, maximum diameter, average density or intensity, shape index, etc. These quantitative indicators have important clinical significance for disease diagnosis, staging, treatment effect evaluation, and prognosis.

[0037] Uncertainty Quantification: Combining techniques such as Bayesian deep learning and Monte Carlo dropout, the model's prediction uncertainty is quantified, providing model confidence assessment for doctors, which helps risk management and decision support.

[0038] Report Generation and Visualization: Structured Report Generation: The final analysis results are integrated into a structured report template, including detailed information of lesions, quantitative measurement results, imaging feature descriptions, and possible diagnostic suggestions. The report can automatically fill in predefined fields and allow doctors to revise and supplement.

[0039] Result Visualization: The segmentation mask or bounding box of the lesion is overlaid on the original medical image, and the lesion area is highlighted in the form of pseudo-color map, etc. Three-dimensional reconstruction view is provided to help doctors intuitively understand the spatial position and morphology of the lesion, enhancing the explainability of diagnosis.

[0040] Trend Analysis: For follow-up images, the system can compare the quantitative indicators of the same lesion at different time points, generate lesion change trend chart, and assist doctors in evaluating treatment effect or disease progression.

[0041] Clinical Decision Support: The analysis results are integrated with the patient's clinical history, laboratory examination results and other information to provide more comprehensive clinical decision support for doctors. The system can also provide recommendations of relevant medical literature or guidelines to assist doctors in making more accurate diagnosis and treatment plan selection.

[0042] The deep learning-based medical image computer-aided analysis method of the present application, through the synergistic effect of the above specific steps, provides a comprehensive, efficient, accurate and robust solution. It effectively improves the automation level and objectivity of medical image analysis, reduces the workload of doctors, and reduces the risk of missed diagnosis and misdiagnosis. It has far-reaching significance for accelerating medical research progress and improving patient medical services. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is the overall technical scheme architecture schematic diagram of the deep learning-based medical image computer-aided analysis method proposed by the present application; Figure 2 is the core principle framework schematic diagram of the lightweight multi-scale feature extraction network in the present application. DETAILED DESCRIPTION

[0044] In the following description, specific terminology and structural descriptions are used to set forth the embodiments of the present application for the convenience of understanding. However, these descriptions are not intended to limit the present application, and those skilled in the art can understand that various modifications and substitutions can be made without departing from the spirit and scope of the present application. The accompanying drawings are merely used to illustrate the exemplary embodiments of the present application, are not drawn to scale, and should not be interpreted as limiting the present application.

[0045] The present application proposes a deep learning-based medical image computer-aided analysis method. This method improves the recognition ability of complex lesions and microstructures in medical images by using deep learning technology, especially by introducing a lightweight multi-scale feature extraction network. Specifically, this method can efficiently extract multi-level and multi-scale semantic features from original medical images and perform accurate medical image analysis, such as lesion segmentation, disease classification, or anomaly detection, on this basis. Through the adaptive learning ability of the deep learning model, this method can effectively handle the heterogeneity of image data, reduce the dependence on manual interpretation, and improve the automation and consistency of analysis. Ultimately, this method aims to provide a high-precision, high-efficiency, and high-generalization medical image computer-aided analysis tool, thereby providing objective and reliable diagnostic evidence for clinicians and assisting in the early detection and precise treatment of diseases.

[0046] As shown in Figure 1 and Figure 2 , the deep learning-based medical image computer-aided analysis method according to the embodiments of the present application includes: S110, acquiring medical image data and performing preprocessing; S120, performing feature extraction on the preprocessed medical image data through a lightweight multi-scale feature extraction network; S130, performing medical image analysis based on the extracted features; and S140, outputting the analysis results and generating a report.

[0047] In the above deep learning-based medical image computer-aided analysis method, the step S110, the medical image data is acquired and preprocessed. It should be understood that acquiring and preprocessing the medical image data is the basis for ensuring that the subsequent deep learning model can efficiently and accurately analyze. Medical image data usually comes from multiple modalities, including but not limited to computed tomography (slicing imaging), magnetic resonance imaging (magnetic resonance), positron emission tomography (positron emission computed tomography), ultrasonic examination, and pathological section images. These raw data have significant differences in format, resolution, contrast, noise level, etc. Directly inputting the deep learning model will seriously affect the training efficiency and generalization performance of the model. Specifically, the data acquisition stage involves collecting image files from the picture archiving and communication system (picture archiving and communication system), digital imaging and communications in medicine (digital imaging and communications in medicine) compatible devices or special databases. These files are usually stored in digital imaging and communications in medicine format, containing image pixel information, patient information, scanning parameters and image metadata. The acquired data first needs to be converted and unified to ensure that all images can be recognized and loaded by the subsequent processing module. For example, multiple digital imaging and communications in medicine files generated by different manufacturers' devices are uniformly parsed into internal standard data structures for unified management and access.

[0048] Further, after acquiring the medical image data, the preprocessing process aims to eliminate redundant information of the original data, correct noise, unify the data scale, and enhance the contrast of the target lesion. The preprocessing specifically includes the following operations: first, data standardization and normalization. The image intensity value range produced by different image devices may differ greatly, and direct training will cause the model to converge difficultly. The standardization operation is to adjust the pixel intensity value to the distribution of zero mean and unit variance, while the normalization operation scales the pixel intensity value to a specific numerical interval, such as zero to one or minus one to one. This process helps to stabilize model training and improve the adaptability of the model to different data sources. Second, image registration. For multi-modal images or images of the same organ collected at different time points, image registration technology is needed to align them to the same coordinate system to ensure the spatial consistency of the corresponding anatomical structures. Image registration uses feature-based methods or mutual information-based methods to achieve geometric transformation of images by optimizing similarity measures. This process is crucial for fusing multi-modal information or tracking disease progression. Third, noise suppression and artifact removal. Medical images often contain noise and artifacts caused by scanning devices, patient motion, or environmental factors. For example, the beam hardening artifact in tomographic images or the motion artifact in magnetic resonance images. The preprocessing stage uses various filtering techniques, such as Gaussian filtering, median filtering, or non-local mean filtering, as well as deep learning-based denoising algorithms to improve image quality and lesion visibility. Fourth, contrast enhancement. Through histogram equalization, gamma correction or contrast stretching techniques based on domain knowledge, the contrast between the target lesion and the surrounding tissue is effectively improved, making it more prominent in the image, which is convenient for subsequent feature extraction. Fifth, data augmentation. In order to increase the diversity and generalization ability of the training data set, the preprocessing process also includes data augmentation operations. This operation generates new training samples by performing geometric and photometric transformations such as random rotation, flipping, scaling, cropping, brightness adjustment, and elastic deformation on existing images, effectively reducing the risk of model overfitting and improving the robustness of the model to various variations. The fine implementation of the above preprocessing steps provides high-quality input for the lightweight multi-scale feature extraction network, ensuring the accuracy and reliability of subsequent analysis.

[0049] In the above deep learning-based medical image computer-aided analysis method, the step S120 performs feature extraction on the preprocessed medical image data by a lightweight multi-scale feature extraction network. It should be understood that the size, shape and position of the lesions in the medical image are greatly different, and the boundaries thereof are often blurred. Traditional feature extraction methods, such as hand-designed texture features or edge features, are difficult to effectively capture these multi-scale and multi-level semantic information, resulting in insufficient performance in identifying small lesions or complex anatomical structures. If a single scale convolution kernel is used for feature extraction, it may result in loss of local details or lack of global context information. At the same time, in order to meet the requirements of clinical application on computing efficiency and deployment flexibility, the feature extraction network must have the characteristics of lightweight, avoiding excessive model parameters and computational complexity. Therefore, the present method introduces a lightweight multi-scale feature extraction network, aiming to balance the comprehensiveness of feature extraction and the economy of computing resources.

[0050] Specifically, the design principle of the present lightweight multi-scale feature extraction network is as follows: first, a depth separable convolution is used. The depth separable convolution decomposes the standard convolution operation into two independent steps of depth convolution and point convolution, greatly reducing the model parameter amount and computational complexity, while maintaining similar feature extraction capability. The depth convolution performs convolution on each input channel independently, effectively extracting spatial features; the point convolution integrates information between different channels through one-by-one convolution. This design enables the network to effectively capture local texture and structure information of the image under the premise of lightweight. Second, multi-scale features are fused. The network contains multiple parallel or serial branches, each branch using convolution kernels with different receptive fields or different down-sampling rates to process the input image, thereby capturing feature information of different scales. For example, by setting a convolution kernel with different hole rates, the receptive field is effectively expanded without increasing the parameter and computational complexity, and more extensive context information is obtained. The fusion of multi-scale features is realized through a feature pyramid network (pyramid feature network) structure, which fuses the high-resolution detailed features of the low layer with the semantically rich features of the high layer, forming a feature representation containing rich details and global context information.

[0051] Third, attention mechanism is introduced. To enable the network to adaptively focus on the most meaningful areas in medical images (e.g. lesion areas) for diagnosis and suppress irrelevant background noise, the lightweight multi-scale feature extraction network integrates channel attention mechanism and spatial attention mechanism. Channel attention mechanism learns the importance weight of each channel feature to weight different semantic information; spatial attention mechanism learns the importance weight of each spatial position to enhance the feature representation of key areas. This mechanism enables the network to dynamically adjust its attention to different areas and different features, further improving the accuracy of lesion recognition. Fourth, efficient activation function and normalization layer are used. To maintain the lightweight nature of the network, high-efficiency and high-performance activation functions such as leaky rectified linear unit (leaky rectified linear unit) or hybrid attention activation function are selected. At the same time, batch normalization layer (batch normalization) or instance normalization layer (instance normalization) is used to accelerate model training, improve model stability, and reduce internal covariate shift.

[0052] Through the above design, the lightweight multi-scale feature extraction network can effectively solve the shortcomings of traditional methods in multi-scale lesion recognition and computational efficiency. The output feature representation is a high-dimensional semantic vector that contains rich anatomical structure information and pathological features in medical images. These features not only encode local information such as shape, size, and texture of lesions, but also integrate global context information such as spatial relationship with surrounding tissues. These feature vectors will be input into the subsequent medical image analysis module as the basis for accurate diagnosis.

[0053] In the embodiment of the present application, the step S120 of feature extraction on the preprocessed medical image data, the lightweight multi-scale feature extraction network optimizes the model parameters by minimizing the difference between the prediction results and the true labels in the training stage. This process involves a loss function to quantify model performance. Common loss functions include a combination of cross-entropy loss (cross-entropy loss) and Jaccard coefficient (Jaccard index) loss, which can be expressed as: wherein, represents the total loss function, represents the cross-entropy loss, represents the Jaccard coefficient loss, and are the weight coefficients of the two loss terms. The cross-entropy loss is calculated as follows: wherein, represents the total number of pixels, represents the pixel the true label of each pixel (e.g., zero for background and one for lesion), pixel points predicted by the model the probability that each pixel belongs to a lesion. The Jaccard coefficient loss The calculation formula is as follows: The combination of the above loss functions enables the model to focus on the classification accuracy at the pixel level and consider the overall structure and boundary matching degree of the lesion area during the optimization process, thereby achieving more accurate lesion recognition and segmentation in complex medical image analysis tasks.

[0054] In the above deep learning-based medical image computer-aided analysis method, the step S130 performs medical image analysis based on the extracted features. It should be understood that based on the rich semantic features obtained by the lightweight multi-scale feature extraction network in step S120, this step aims to perform specific medical image analysis tasks such as lesion segmentation, disease classification, or anomaly detection. These tasks are key links in clinical diagnosis and directly affect the early detection, accurate positioning, and treatment plan of diseases. Specifically, the extracted features encode various information in medical images from low-level textures, edges to high-level semantics (e.g., organ structure, lesion morphology). How to effectively utilize these features for accurate downstream task analysis is the core of this step.

[0055] For the lesion segmentation task, the method inputs the extracted features into a dedicated segmentation head (segmentation head). This segmentation head usually includes a series of up-sampling layers, convolutional layers, and activation layers, aiming to gradually restore the low-resolution semantic feature map to the resolution of the original image and predict the class it belongs to at each pixel level (e.g., normal tissue, tumor, inflammation). The up-sampling operation is achieved through transposed convolution (deconvolution) or bilinear interpolation to increase the spatial dimension of the feature map. The convolutional layer is used to further refine the feature representation and fuse information from different scales. Finally, an output layer (e.g., a convolutional layer with a nonlinear activation function) generates a probability map for each pixel belonging to a specific class. For example, in a tumor segmentation task, the model outputs a probability map with the same size as the original image, where each pixel value represents the probability that the pixel belongs to the tumor region. Then, by setting a threshold, a binary segmentation mask can be obtained, clearly indicating the precise location and boundary of the lesion. The key to this process is that the segmentation head must be able to fully utilize the multi-scale feature information provided by the lightweight multi-scale feature extraction network to ensure the accuracy and detail integrity of the segmentation results.

[0056] For disease classification or abnormality detection tasks, the extracted features are passed through a classification head. This classification head is usually composed of one or multiple fully connected layers, which are used to map the high-dimensional feature vectors to the final class probabilities. Batch normalization layers and activation functions can be inserted between the fully connected layers to enhance the non-linear expressiveness of the model and speed up the training. The input of the classification head is the global feature representation output by the lightweight multi-scale feature extraction network, e.g., the vector obtained by reducing the dimensionality of the feature map through global average pooling or global max pooling. Finally, through an output layer with a non-linear activation function (e.g., a non-linear activation function for binary classification or a soft-max function for multi-classification), the probabilities of the image belonging to each disease class or normal / abnormal class are generated. By thresholding these probabilities, the final diagnosis result can be obtained.

[0057] It should be understood that the effectiveness of this step lies in the powerful pattern recognition ability of the deep learning model, which can automatically learn the complex mapping relationship from medical image features to clinical diagnosis results, thereby avoiding the tediousness and subjectivity of manual feature extraction in traditional methods and significantly improving the objectivity and accuracy of the analysis. At the same time, by designing corresponding analysis modules for different tasks (segmentation, classification), the generality and flexibility of the method are ensured, which can adapt to different clinical needs.

[0058] In the above deep learning-based medical image computer-aided analysis method, the step S140 outputs the analysis results and generates a report. It should be understood that presenting the analysis results of the deep learning model in an intuitive, easy-to-understand, and clinically standardized manner is a key step in realizing the value of computer-aided analysis. Pure model prediction results, such as probability values or segmentation masks, need further interpretation and transformation to assist in decision-making for clinicians. Therefore, this step aims to convert the original output of the model into a structured report with clinical guidance significance and provide visualized result display.

[0059] Specifically, the output analysis results include but are not limited to the following forms: first, quantitative indicators. For lesion segmentation tasks, the output results include the volume, size (length, width, height), average density value or intensity value, and shape features (e.g., sphericity, flatness) of the lesion. These quantitative indicators can provide objective descriptions of the lesion for doctors, assisting in evaluating the severity and progression of the disease. For classification tasks, the output results are the probability values of each class, such as the probability of a tumor being benign or malignant. These probability values can serve as a reference for doctors when making diagnoses.

[0060] Second, visualization results. By superimposing the segmentation mask onto the original medical image, the precise location and extent of the lesion can be visually demonstrated, such as highlighting the tumor area, inflammation area or organ boundary with different colors. For classification tasks, a heat map (heat map) can be labeled on the key area of the image to visualize the key area where the model makes a judgment, such as through the gradient weighted class activation mapping (gradient weighted class activation mapping) technique, to enhance the explainability of the model results. These visualization results can help doctors quickly understand the analysis basis of the model and conduct manual review.

[0061] Third, structured report generation. According to the pre-defined report template, this method automatically generates a structured report containing all key information. The report content usually includes: patient basic information, image examination type and date, model analysis results (quantitative indicators, classification probability), visualization images (images superimposed with segmentation results, heat maps), model confidence score, and potential clinical recommendations. The format of the report is compatible with the hospital information system (hospital information system) or the radiology information system (radiology information system), which facilitates seamless integration into the existing clinical workflow. For example, a table containing tumor volume, boundary clarity, internal uniformity and other parameters is automatically generated in the report, along with a multi-view lesion 3D reconstruction image. This structured report not only improves the efficiency of report generation, but also ensures the consistency and completeness of information presentation.

[0062] Fourth, exception handling and quality control. In the result output stage, the system also integrates an exception handling mechanism and a quality control module. This module can check the rationality of the model's prediction results, such as checking whether the segmentation results are connected and whether the classification probability is within a reasonable range. If an exception is found, the system will issue an alert and may trigger a manual review process to ensure the accuracy and reliability of the final output report. For example, if the model predicts a lesion volume that is too large or too small, beyond the normal range, the system will prompt the doctor for a second confirmation. This mechanism effectively improves the safety and robustness of the entire computer-aided analysis system.

[0063] In summary, the deep learning-based medical image computer-aided analysis method based on the embodiments of the present application is illustrated, which uses deep learning technology to introduce a lightweight multi-scale feature extraction network, which can efficiently extract multi-level and multi-scale semantic features from the original medical image, and on this basis, perform accurate medical image analysis, such as lesion segmentation, disease classification or anomaly detection, and finally output in the form of structured report and visualization results, thereby providing objective and reliable diagnostic basis for clinicians, assisting in early detection and precise treatment of diseases.

Claims

1. A medical image computer-aided analysis method based on deep learning, characterized in that: The following steps are involved: receiving medical image data to be analyzed, the medical image data including single-modality images or multi-modality images, and performing metadata parsing and image sequence matching; Performing preprocessing operations on the medical image data to eliminate noise and artifacts in the original data and unify the image format and pixel values; By constructing a lightweight multi-scale feature extraction network, deep semantic information and multi-scale image feature representation of the medical image data are obtained; Based on the multi-scale image feature representation and pre-trained deep learning model, a fine-grained lesion recognition module is used to perform lesion recognition or segmentation. During the recognition process, an adaptive regularization strategy is used to enhance the generalization ability of the model. A multimodal deep fusion mechanism is applied to fully utilize the complementary information of different modal images to generate preliminary analysis results. The preliminary analysis results are post-processed and evaluated to output a final computer-assisted analysis report containing quantitative indicators, visualization content, and a structured report.

2. The medical image computer-aided analysis method based on deep learning according to claim 1, characterized in that: Receiving the medical imaging data to be analyzed specifically includes: Obtaining original medical images from medical imaging devices or data storage systems; The medical imaging data includes at least one of computed tomography images, magnetic resonance imaging images, positron emission tomography images, ultrasound images, digital subtraction angiography images, and pathological section images; Perform metadata parsing to obtain image acquisition parameters, patient information, and scan protocols; For multimodal images, image sequences are matched and associated to ensure that image data of different modalities correspond to the same lesion area of ​​the same patient.

3. The medical image computer-aided analysis method based on deep learning according to claim 1, characterized in that: The preprocessing operation on the medical image data specifically includes: Image format conversion and data normalization: convert the received image data in different formats into a standard format and normalize the pixel values ​​of the image; Image registration: For multimodal image data, accurate registration operations are performed to spatially align images of different modalities; Noise suppression and artifact removal, using filtering algorithms to suppress random noise in images and identify and remove artifacts caused by device and patient motion or the reconstruction process; Region of interest cropping and enhancement: Based on prior knowledge or automatic detection algorithms, the region of interest containing potential lesions is cropped and local contrast enhancement is performed on the region of interest.

4. The method for computer-assisted analysis of medical images based on deep learning according to claim 3, characterized in that: The image registration adopts an intensity-based or feature-based algorithm combined with affine transformation or elastic transformation for geometric correction; the noise suppression and artifact removal adopt Gaussian filtering, median filtering or non-local mean filtering; the region of interest cropping and enhancement adopts histogram equalization or adaptive histogram equalization.

5. The medical image computer-aided analysis method based on deep learning according to claim 1, characterized in that: The lightweight multi-scale feature extraction network includes: The encoder-decoder architecture consists of an encoder part that extracts high-level semantic features through convolutional layers and pooling layers, and a decoder part that restores the spatial resolution of the feature map through upsampling layers and skip connections; Multi-scale feature fusion module, which captures contextual information of different scales through spatial pyramid pooling module, feature pyramid network structure or void convolution; Lightweight convolution modules that use depthwise separable convolution, group convolution, or channel attention mechanisms to reduce computational complexity and memory consumption; Adaptive attention mechanism, which introduces a channel attention module or a spatial attention module to dynamically focus on key diagnostic areas in the image; The feature dimension compression and enhancement module uses a bottleneck structure or residual connection to perform feature dimension compression and enhancement at different levels.

6. The method for computer-assisted analysis of medical images based on deep learning according to claim 5, characterized in that: During the training phase, the lightweight multi-scale feature extraction network optimizes model parameters by minimizing the difference between the predicted results and the true labels. The prediction result minimization process uses a combination of cross entropy loss and Jaccard coefficient loss as the loss function.

7. The method for computer-assisted analysis of medical images based on deep learning according to claim 1, characterized in that: Performing lesion identification or segmentation based on the multi-scale image feature representation and the pre-trained deep learning model specifically includes: Inputting the multi-scale image feature representation into an object detection model, wherein the model predicts the category and bounding box coordinates of the lesion based on a single-stage detector or a two-stage detector to perform lesion detection and localization; Inputting the multi-scale image feature representation into a semantic segmentation model or an instance segmentation model, wherein the model restores the spatial resolution through a decoder and combines skip connections to generate a pixel-level lesion mask to segment the lesion area; The multi-scale image feature representation is input into a classification model, and the model classifies and identifies lesions by learning the morphological, textural and functional features of the lesions.

8. The medical image computer-aided analysis method based on deep learning according to claim 1, characterized in that: The multimodal deep fusion mechanism is carried out in at least one of the following stages: Early fusion: In the feature extraction stage, the original images or their preliminary features of different modalities are spliced ​​or fused at the element level, and then input into a unified deep learning model for processing; Mid-term fusion, interactively fusing features extracted from different modalities at the intermediate level of the network, using a cross-modal attention mechanism or learning shared representations through a common embedding space; Late fusion: images of different modalities are analyzed independently, and then the respective recognition or segmentation results are fused and decided through a voting mechanism, weighted average or meta-learning model.

9. The medical image computer-aided analysis method based on deep learning according to claim 1, characterized in that: The adaptive regularization strategy includes at least one of the following techniques: Data augmentation: randomly flipping, rotating, scaling, cropping, adjusting brightness, or elastically deforming training data; Batch normalization, adding a batch normalization layer after the network layer; Dropout, randomly discarding the output of some neurons during training; Weight decay, adding a penalty term for model weights to the loss function; Label smoothing, which softens hard labels in classification tasks; Contrastive learning uses unsupervised or self-supervised methods to learn discriminative feature representations of images.

10. The medical image computer-aided analysis method based on deep learning according to claim 1, characterized in that: Post-processing and result evaluation of the preliminary analysis results to output a final computer-assisted analysis report specifically includes: Performing connected domain analysis, morphological operations and false positive reduction on the preliminary analysis results; Perform precise quantitative measurements of identified or segmented lesions, including volume, maximum diameter, average density or intensity, and shape index; Combined with Bayesian deep learning or Monte Carlo dropout techniques to quantify the uncertainty of model predictions; Generate structured reports containing detailed information about the lesion, quantitative measurements, descriptions of imaging features, and diagnostic recommendations; Result visualization: superimpose the segmentation mask or bounding box of the lesion on the original medical image, highlight the lesion area in the form of pseudo-color image, and provide a three-dimensional reconstructed view; Perform trend analysis, compare the quantitative indicators of the same lesion at different time points, and generate a lesion change trend graph.

Images