Medical image automatic positioning system based on deep learning

Abstract

The invention discloses a medical image automatic positioning system based on deep learning, which comprises an image preprocessing module, a feature extraction module, a positioning optimization module and a post-processing module, and is characterized in that multi-level spatial features are extracted by constructing a five-layer multi-scale feature pyramid network, and accurate positioning is realized by adopting a three-level cascade attention mechanism; and finally, a positioning result is optimized through a self-adaptive post-processing algorithm. According to the invention, a space and channel dual attention mechanism is innovatively designed, so that the positioning precision Dice coefficient reaches 0.92 + / -0.03, and the positioning precision is improved by more than 40%. The system is particularly suitable for positioning key parts such as tumors and vasculopathy, and reliable technical support is provided for precise medical treatment.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and more specifically to a deep learning-based automatic medical image localization system. Background Technology

[0002] Currently, medical image localization mainly relies on doctors' professional knowledge and experience. However, with the rapid growth of medical image data, traditional manual localization methods can no longer meet the needs of efficient and accurate clinical practice. In recent years, artificial intelligence technology, especially deep learning technology, has made significant progress in the field of medical image analysis. However, most existing medical image automatic localization systems are based on single-modality image data, such as processing only CT or MRI images, ignoring the complementarity between multimodal image data, which limits the accuracy and reliability of localization; in addition, traditional systems have insufficient localization accuracy at the edge of lesions, making it difficult to meet the needs of precision medicine. Therefore, designing an automated medical image localization system based on deep learning is of great significance. Summary of the Invention

[0003] The purpose of this invention is to provide a deep learning-based automatic medical image localization system to solve the technical problems existing in current medical image localization technologies, such as insufficient localization accuracy, difficulty in multimodal data fusion, and inaccurate localization of lesion edges.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based automatic medical image localization system, comprising: Image preprocessing module: performs standardization and feature enhancement on the input medical images, and extracts key feature information from multimodal medical images; Feature extraction module: Connected to the image preprocessing module, it constructs a multi-scale feature pyramid network to extract multi-level spatial features from medical images; Localization optimization module: Connected to the feature extraction module, it uses a cascaded attention mechanism to achieve precise localization of the lesion area; Post-processing module: Connected to the positioning optimization module, it performs morphological optimization on the positioning results and outputs the final positioning information.

[0005] In a preferred embodiment, the image preprocessing module includes: Acquire multimodal medical image data, including CT images, MRI images, and ultrasound images; Image standardization processes include grayscale normalization and spatial normalization; Data augmentation techniques are used to expand the training samples, including elastic deformation and random cropping.

[0006] In a preferred embodiment, the feature extraction module includes: Construct a multi-scale feature pyramid network, including a low-level detailed feature extraction layer, a mid-level spatial feature fusion layer, and a high-level semantic feature extraction layer; The low-level detail feature extraction layer uses a 3×3 convolutional kernel to extract local detail features; The intermediate spatial feature fusion layer integrates spatial feature information at different scales; A high-level semantic feature extraction layer obtains global semantic information of the image.

[0007] In a preferred embodiment, the positioning optimization module includes: A cascaded attention mechanism is adopted, including a spatial attention submodule and a channel attention submodule; The spatial attention submodule enhances the spatial feature representation of the lesion region; The channel attention submodule optimizes the weight allocation of feature channels; Positioning accuracy is gradually optimized through cascading.

[0008] In a preferred embodiment, the post-processing module includes: Perform morphological closing operations on the localization results to eliminate small voids; An adaptive edge optimization algorithm is used to smoothly locate the boundary; Output the final localization result, including the bounding box coordinates and confidence score.

[0009] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs a multi-scale feature pyramid network, which can simultaneously capture local detail features and global semantic information of medical images, solving the problem of inaccurate lesion edge localization in traditional systems. The network employs a five-layer structure: the bottom layer extracts fine texture features using 3×3 convolutional kernels; the middle layer fuses multi-scale information through a feature pyramid structure; and the top layer captures semantic features using global average pooling. This multi-layer feature extraction method improves localization accuracy by more than 40%, particularly excelling in the localization of critical areas such as tumors and vascular lesions.

[0010] 2. This invention innovatively designs a three-level cascaded attention mechanism, which significantly improves the localization accuracy of the lesion region through the synergistic effect of spatial and channel attention mechanisms. The spatial attention submodule uses a 7×7 convolutional kernel to generate a spatial weight map, while the channel attention submodule optimizes channel features through a fully connected layer.

[0011] 3. The adaptive post-processing algorithm proposed in this invention, combining morphological operations and gradient information, effectively eliminates localization noise, resulting in smoother and more accurate localization boundaries. This algorithm preserves more details in areas with high edge intensity and performs stronger smoothing in smooth areas. The final localization output includes multi-dimensional information such as three-dimensional spatial coordinates, bounding box size, and localization reliability score, possessing extremely high clinical practical value. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0013] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Example 1, please refer to Figure 1 As shown in this embodiment, the deep learning-based automatic medical image localization system includes: Image preprocessing module: performs standardization and feature enhancement on the input medical images, and extracts key feature information from multimodal medical images; Feature extraction module: Connected to the image preprocessing module, it constructs a multi-scale feature pyramid network to extract multi-level spatial features from medical images; Localization optimization module: Connected to the feature extraction module, it uses a cascaded attention mechanism to achieve precise localization of the lesion area; Post-processing module: Connected to the positioning optimization module, it performs morphological optimization on the positioning results and outputs the final positioning information; To further explain, currently, medical image localization mainly relies on doctors' professional knowledge and experience. However, with the rapid growth of medical image data, traditional manual localization methods can no longer meet the needs of efficient and accurate clinical practice. In recent years, artificial intelligence technology, especially deep learning technology, has made significant progress in the field of medical image analysis. However, most existing medical image automatic localization systems are based on single-modality image data, such as processing only CT or MRI images, ignoring the complementarity between multimodal image data, which limits the accuracy and reliability of localization; in addition, traditional systems have insufficient localization accuracy at the edge of lesions, making it difficult to meet the needs of precision medicine. This invention constructs a multi-scale feature pyramid network, enabling simultaneous capture of local detail features and global semantic information in medical images, thus solving the problem of inaccurate lesion edge localization in traditional systems. Employing a cascaded attention mechanism, the synergistic effect of spatial and channel attention submodules significantly improves the localization accuracy of lesion regions. Finally, a morphological optimization algorithm effectively eliminates noise interference in the localization results, yielding smoother and more accurate localization boundaries. Compared with existing technologies, this invention demonstrates higher accuracy and robustness in multimodal medical image localization tasks, providing more reliable technical support for clinical diagnosis and treatment.

[0016] In one embodiment, the image preprocessing module includes: Acquire multimodal medical image data, including CT images, MRI images, and ultrasound images; Image standardization processes include grayscale normalization and spatial normalization; Data augmentation techniques are used to expand the training samples, including elastic deformation and random cropping; To further elaborate, during the data acquisition phase, we established a multi-center medical image database, collecting medical image data from three modalities: CT, MRI, and ultrasound. Each modality contained at least 1000 labeled samples. Image standardization involved the following steps: first, grayscale normalization was performed on each image, mapping pixel values ​​to the 0-1 range; then, spatial standardization was performed, resampling all images to a uniform resolution of 1mm × 1mm × 1mm. For data augmentation, we employed elastic deformation technology to simulate the natural deformation of organs and tissues, setting the deformation coefficient α = 100 and the standard deviation σ = 10 using a Gaussian kernel. Random cropping used a fixed size of 256 × 256 pixels to ensure spatial consistency across different modalities. This preprocessing workflow significantly improved the subsequent feature extraction results and enhanced the model's generalization ability.

[0017] In one embodiment, the feature extraction module includes: Construct a multi-scale feature pyramid network, including a low-level detailed feature extraction layer, a mid-level spatial feature fusion layer, and a high-level semantic feature extraction layer; The low-level detail feature extraction layer uses a 3×3 convolutional kernel to extract local detail features; The intermediate spatial feature fusion layer integrates spatial feature information at different scales; A high-level semantic feature extraction layer obtains global semantic information of the image; To further explain, the specific architecture of the multi-scale feature pyramid network is as follows: The bottom-level detail feature extraction layer consists of five consecutive 3×3 convolutional blocks, each containing a convolutional layer, a batch normalization layer, and a ReLU activation function, with a stride of 1 and padding to maintain the feature map size. The middle-level spatial feature fusion layer adopts a feature pyramid structure, fusing feature maps of different scales (1 / 4, 1 / 8, and 1 / 16 of the original size) through top-down and lateral connections. After adjusting the number of channels using 1×1 convolutions, it performs upsampling through bilinear interpolation. The high-level semantic feature extraction layer uses global average pooling combined with fully connected layers to extract global contextual information of the image. After pre-training on the ImageNet dataset, the network was fine-tuned on medical image data. The final output multi-scale feature map contains rich information from local to global perspectives, laying a solid foundation for subsequent accurate localization.

[0018] In one embodiment, the positioning optimization module includes: A cascaded attention mechanism is adopted, including a spatial attention submodule and a channel attention submodule; The spatial attention submodule enhances the spatial feature representation of the lesion region; The channel attention submodule optimizes the weight allocation of feature channels; Positioning accuracy is gradually optimized through cascading. To further explain, the cascaded attention mechanism is implemented as follows: the spatial attention submodule first performs global average pooling and max pooling on the input feature map, concatenates the results, and generates a spatial attention weight map through a 7×7 convolutional layer; the channel attention submodule obtains the statistical features of each channel through global average pooling, generates channel attention weights through two fully connected layers (the intermediate layer dimension is C / 16, where C is the number of input channels) and a sigmoid activation function; the outputs of the two submodules are multiplied together and then multiplied element-wise with the original feature map to achieve feature recalibration. This embodiment adopts a three-level cascaded structure, with the output of each level of attention module serving as the input of the next level. Through this progressive optimization strategy, the final lesion region localization accuracy reaches sub-millimeter level (0.5mm), significantly better than traditional systems.

[0019] In one embodiment, the post-processing module includes: Perform morphological closing operations on the localization results to eliminate small voids; An adaptive edge optimization algorithm is used to smoothly locate the boundary; Output the final localization result, including the bounding box coordinates and confidence score; Further explanation: The specific technical details of the post-processing stage are as follows: Morphological closing operation uses a 3×3 circular structuring element, performing dilation followed by erosion to effectively fill small holes in the localization result; the adaptive edge optimization algorithm, based on image gradient information, retains more details in areas with high edge intensity and performs stronger smoothing in smooth areas; the final output localization result includes the following information: 1) three-dimensional spatial coordinates (x, y, z) of the lesion area; 2) bounding box dimensions (length, width, height); 3) localization reliability score (0-1 interval). Clinically verified, the Dice coefficient of the localization result of this invention compared with the expert manually annotated result reaches 0.92±0.03, demonstrating extremely high clinical practical value.

[0020] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based automatic medical image localization system, characterized in that: include: Image preprocessing module: performs standardization and feature enhancement on the input medical images; Feature extraction module: Constructs a multi-scale feature pyramid network to extract multi-level spatial features; Location optimization module: Employs a cascaded attention mechanism to achieve precise positioning; Post-processing module: Performs morphological optimization on the localization results.

2. The deep learning-based automatic medical image localization system according to claim 1, characterized in that: The image preprocessing module includes: Acquire multimodal medical image data and perform standardization processing. The multimodal medical image data includes CT images, MRI images, and ultrasound images. Data augmentation techniques are used to expand the training samples, including elastic deformation and random cropping.

3. The deep learning-based automatic medical image localization system according to claim 1, characterized in that: The feature extraction module includes: Construct a five-layer multi-scale feature pyramid network, wherein: The bottom layer uses 3×3 convolution kernels to extract local texture features; The middle layer fuses multi-scale information through a feature pyramid structure; The higher-level layer utilizes global average pooling to capture semantic features.

4. The deep learning-based automatic medical image localization system according to claim 1, characterized in that: The positioning optimization module includes: A three-level cascaded attention mechanism is adopted, including: Spatial attention submodule: Uses 7×7 convolutional kernels to generate spatial weight maps; Channel attention submodule: Optimizes channel features through a fully connected layer.

5. The deep learning-based automatic medical image localization system according to claim 1, characterized in that: The post-processing module includes: Morphological closing operations are performed on the localization results to eliminate small voids; An adaptive edge optimization algorithm based on image gradients is used to smoothly locate the boundary. The output includes the localization results, which include 3D coordinates, bounding box dimensions, and localization reliability score.

6. The deep learning-based automatic medical image localization system according to claim 4, characterized in that: The Dice coefficient of the cascaded attention mechanism reached 0.92±0.

03.

7. The deep learning-based automatic medical image localization system according to claim 1, characterized in that: The system is suitable for locating key areas such as tumors and vascular lesions.

Images