Obstetrics and gynecology image analysis system based on deep learning

Abstract

The invention, which relates to the technical field of medical image analysis, discloses a deep learning-based image analysis system for the obstetrics and gynecology department, comprising a feature extraction module, a region division module, a feature analysis module, an anomaly evaluation module and a visual output module. Multi-scale feature extraction is carried out on medical images of the obstetrics and gynecology department through a multi-resolution convolutional neural network, multi-scale features are integrated by adopting a cross-resolution fusion operation and an adaptive resolution adjustment operator, and accurate region division is carried out on the uterus, the ovary and the fallopian tube through semantic segmentation and morphological screening rules. An improved parallel attention structure is introduced to enhance the feature description capability, an abnormal index is generated based on a local feature quantity and discreteness measurement, and a visual analysis result is generated through color mapping, thermodynamic diagram superposition and a highlight display technology. According to the method, the detection rate and the diagnosis precision of the tiny lesions in the images of the obstetrics and gynecology department are effectively improved, and end-to-end intelligent analysis from feature extraction to visual output is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image analysis, and in particular to a gynecological image analysis system based on deep learning. BACKGROUND

[0002] In recent years, deep learning technology has shown great potential in the field of medical image analysis, especially in the automatic analysis of gynecological ultrasound images.

[0003] At present, in the Chinese invention patent with the publication number CN119067859A, an enhanced analysis method for gynecological medical images is disclosed, which divides the image according to the gray histogram of the to-be-processed uterine B-ultrasound image to obtain a to-be-measured region and performs connected component analysis, determines the initial abnormality degree based on the gray distribution of different connected components and the gray distribution of the to-be-measured region where the connected component is located, determines the distribution complexity of each pixel point according to the gray distribution of the pixel points in the connected component, determines the influence coefficient by combining the distribution complexity and the initial abnormality degree, and determines the to-be-enhanced index of the to-be-measured region; according to the influence coefficient and the to-be-enhanced index, the abnormality index of the connected component is determined; based on the abnormality index, different colors are used to perform color enhancement on each connected component to obtain an enhanced image, but the related technology does not construct a multi-scale feature fusion architecture, which cannot integrate global context information while maintaining high-resolution details, resulting in insufficient perception of small lesions and fuzzy boundaries, lack of adaptive attention mechanism for ultrasound image characteristics, difficulty in accurately focusing on key anatomical structures in complex backgrounds, and the traditional image processing process is disconnected with the deep learning model, and a unified optimization framework cannot be formed to simultaneously improve image quality and diagnostic accuracy. SUMMARY

[0004] The technical problem solved by the present application is that the related technology does not construct a multi-scale feature fusion architecture, which cannot integrate global context information while maintaining high-resolution details, resulting in insufficient perception of small lesions and fuzzy boundaries, lack of adaptive attention mechanism for ultrasound image characteristics, difficulty in accurately focusing on key anatomical structures in complex backgrounds, and the traditional image processing process is disconnected with the deep learning model, and a unified optimization framework cannot be formed to simultaneously improve image quality and diagnostic accuracy.

[0005] To solve the above technical problems, the present application provides the following technical solutions: a gynecological image analysis system based on deep learning, comprising: a feature extraction module, a region division module, a feature analysis module, an abnormality evaluation module and a visualization output module.

[0006] The feature extraction module is used to perform multi-scale feature extraction on the original gynecological medical image through a multi-resolution convolutional neural network to obtain a multi-scale feature map.

[0007] The region division module is configured to divide the original gynecological medical image based on the multi-scale feature map to obtain a divided region.

[0008] The feature analysis module is configured to perform quantitative feature analysis on the divided region to obtain a feature description.

[0009] The anomaly evaluation module is configured to process the feature description to obtain an anomaly index.

[0010] The visual output module is configured to visually mark the divided region based on the anomaly index to obtain an analysis result image.

[0011] As a preferred scheme of the gynecological image analysis system based on deep learning, the processing logic for performing multi-scale feature extraction on the original gynecological medical image through the multi-resolution convolutional neural network includes:

[0012] The original gynecological medical image is input into the multi-resolution convolutional neural network, which is composed of multiple parallel branches.

[0013] The parallel branch is a parallel path in the multi-resolution convolutional neural network for independently processing a specific resolution feature map.

[0014] The first branch maintains the spatial resolution of the original gynecological medical image. With each additional stage in the multi-resolution convolutional neural network, a new branch is added, and the spatial resolution of the new branch is half of the lowest resolution of the previous stage, and the number of channels is doubled.

[0015] At the end of each stage of the multi-resolution convolutional neural network, the output feature maps of all branches of the stage are subjected to a cross-resolution fusion operation.

[0016] After the cross-resolution fusion operation and the processing of the adaptive resolution adjustment operator in multiple stages of the multi-resolution convolutional neural network, the output feature maps of all branches of the last stage of the multi-resolution convolutional neural network are obtained as the multi-scale feature map and output to the region division module.

[0017] As a preferred scheme of the gynecological image analysis system based on deep learning, the cross-resolution fusion operation specifically includes:

[0018] Each branch of the current stage is sequentially updated as a target branch.

[0019] Each branch of the current stage other than the target branch is defined as a source branch, and the output feature map of each source branch is transformed to the same spatial resolution as the target branch through up-sampling to obtain a transformed feature map aligned with the resolution of the target branch.

[0020] All transformed feature maps are aggregated with the output feature map of the target branch itself, and the output of the aggregation result is taken as the updated feature map of the target branch, and the updated feature map of the target branch is taken as the input of the target branch into the next stage;

[0021] After the cross-resolution fusion operation of the first stage to the third stage of the multi-resolution convolutional neural network, the output feature map of the corresponding branch is input to the adaptive resolution adjustment operator;

[0022] The adaptive resolution adjustment operator generates a set of sampling points for each input feature map through its internal sampling point generator, and performs resampling operation on the input feature map using the set of sampling points to obtain an optimized feature map and update the feature representation of the corresponding branch.

[0023] As a preferred scheme of the deep learning-based gynecological image analysis system, the processing process of the adaptive resolution adjustment operator includes:

[0024] The input feature map is subjected to global average pooling processing to obtain a channel feature vector;

[0025] The channel feature vector is input to a parameter prediction network to obtain a dynamic range factor;

[0026] The dynamic range factor is subjected to scalar multiplication operation with an original offset generated through convolution operation to obtain an adjusted offset;

[0027] The adjusted offset is added to an original grid position through pixel reorganization technology to obtain a set of sampling points;

[0028] The original grid position refers to a normalized coordinate matrix with a regular fixed size same as the spatial size of the input feature map, which defines the initial spatial position of each pixel on the input feature map;

[0029] The input feature map is resampled through a bilinear interpolation method based on the set of sampling points to obtain an optimized feature map.

[0030] As a preferred scheme of the deep learning-based gynecological image analysis system, the parameter prediction network is composed of a fully connected layer and a Sigmoid activation function;

[0031] The fully connected layer is used to map the channel feature vector to an initial dynamic range factor;

[0032] The Sigmoid activation function is used to compress the value of the initial dynamic range factor to the interval of 0 to 1 to obtain the dynamic range factor.

[0033] As a preferred scheme of the deep learning-based gynecological image analysis system, the original gynecological medical image is divided into divided regions based on the multi-scale feature map, and the processing logic includes:

[0034] The multi-scale feature map is cross-scale feature fused based on the feature pyramid network to obtain a unified feature map;

[0035] The unified feature map is pixel-level semantic segmented based on the DeepLabv3+ architecture to obtain a semantic segmentation map;

[0036] Based on the semantic segmentation map, the connected component analysis algorithm is used to aggregate the pixel points belonging to the same category and spatially adjacent to obtain an initial region set;

[0037] The morphological features and statistical features of the initial region set are calculated, and the initial region set is filtered and merged according to a preset filtering rule to obtain the divided region;

[0038] Each divided region includes a category label and corresponding geometric information, the category label includes the uterus, the ovary and the fallopian tube, and the corresponding geometric information includes a pixel set of a spatial position in the original gynecological medical image and a boundary coordinate sequence;

[0039] As a preferred scheme of the deep learning-based gynecological image analysis system, the processing logic for quantitatively analyzing the divided region to obtain a feature description includes:

[0040] When the divided region is quantitatively analyzed, an improved parallel attention structure is introduced;

[0041] The parallel attention structure includes a channel attention unit and a spatial attention unit;

[0042] The channel attention unit is used for multi-scale channel weighting based on multiple channel responses, and the calculation formula is:

[0043]

[0044]

[0045]

[0046]

[0047] wherein, represents the weighted feature value of the i-th channel, represents the global average pooling value of the i-th channel, represents the global average pooling value of the i-th channel, represents the global average pooling value of the i-th channel, represents the global average pooling value of the i-th channel, ​​​​a global maximum pooling value of the channel, a pyramid pooling value of the channel, a weight parameter representing a global average pooling value, a weight parameter representing a global maximum pooling value, a weight parameter representing a pyramid pooling value, a height of the multi-scale feature map, a width of the multi-scale feature map, a horizontal coordinate of the spatial position, a vertical coordinate of the spatial position, a feature value of the channel at the spatial position , a total number of channels, performing a pyramid pooling operation on the multi-scale feature map of the channel, a multi-scale feature map of the channel;

[0048] The calculation method of the multi-scale channel weight is:

[0049] ;

[0050] ;

[0051] ;

[0052] ;

[0053] wherein, a weight of the channel, a weight of global average pooling, a weight of global maximum pooling, a weight of pyramid pooling, an activation function;

[0054] The spatial attention unit generates a spatial weight matrix based on multi-scale convolution;

[0055] The channel weight and the spatial weight are cross-weighted to obtain a dynamic weight distribution, and the calculation formula of the cross-weighting is:

[0056] ;

[0057] ;

[0058] wherein,​ Represents the weight tensor. Indicates spatial weights, This indicates a splicing operation. This indicates the dimension of the concatenated weight tensor. This represents the weight tensor obtained by concatenation. Indicates combined operations;

[0059] The dynamic weight distribution is multiplied element-wise with the input feature map to obtain the enhanced feature representation, which is then used as the feature description.

[0060] As a preferred embodiment of the deep learning-based obstetric and gynecological image analysis system of the present invention, the abnormal indicators are generated based on feature description, and the processing logic includes:

[0061] Local feature quantities are extracted from the feature description, including texture contrast within a region, statistics of gradient magnitude within a region, and feature difference between sub-regions.

[0062] The quantitative features in the feature description are subjected to distribution statistical processing to obtain a set of feature values;

[0063] The discreteness metric parameter is calculated for the feature value set, and the discreteness metric parameter and local feature quantity are calculated by linear weighting method to obtain the intermediate evaluation quantity;

[0064] The intermediate evaluation values ​​are normalized to obtain the initial outlier values;

[0065] Regional statistics are established based on the feature set of the divided regions. Regional weights are obtained by correlating regional statistics with initial outliers.

[0066] Anomaly indices are obtained by linearly combining regional weights and initial anomaly values.

[0067] The anomaly indicators include regional anomaly score, anomaly heat value, regional risk level, and anomaly distribution matrix.

[0068] As a preferred embodiment of the deep learning-based obstetric and gynecological image analysis system of the present invention, the system involves visually labeling the segmented regions based on abnormal indicators to obtain the analysis result image, and its processing logic includes:

[0069] Get each partitioned region;

[0070] Based on the abnormal indicators of each segmented region, display colors are assigned to them according to predefined color mapping rules. In a blank matrix with the same size as the original obstetric and gynecological medical image, color is filled at the corresponding positions in the image according to its geometric information to obtain the region labeling layer.

[0071] According to the geometric information of each divided region, the corresponding abnormal index value is filled to all pixel positions of the region in the image to obtain a spatial abnormal distribution map;

[0072] The spatial abnormal distribution map is subjected to spatial interpolation processing to obtain an abnormal heat map;

[0073] The region marking layer and the abnormal heat map layer are superimposed on the original gynecological medical image in a semi-transparent manner to obtain an analysis result image.

[0074] As a preferred scheme of the gynecological image analysis system based on deep learning provided by the application, the visualization output module further comprises abnormal region highlighting display processing, and the processing logic comprises:

[0075] The region abnormal score in the abnormal index is taken as a judgment basis, and the region abnormal score is compared with a preset high abnormal threshold;

[0076] If the region abnormal score of the divided region is greater than or equal to the preset high abnormal threshold, the region is determined as a high abnormal region;

[0077] If the region abnormal score of the divided region is less than the preset high abnormal threshold, the region is determined as a normal region;

[0078] The high abnormal region is subjected to special marking different from ordinary regions for enhanced display based on the analysis result image;

[0079] The special marking comprises a dynamic flashing effect, a thickened boundary contour line and a superimposed preset warning symbol;

[0080] The analysis result image is subjected to interactive adjustment, and the interactive adjustment comprises zooming, rotating and superimposing different levels of abnormal information.

[0081] The application has the following beneficial effects: through the multi-resolution convolutional neural network and the cross-resolution fusion operation, efficient integration of multi-scale features is realized, the perception ability for micro lesions and fuzzy boundaries is enhanced while high-resolution details are maintained; the adaptive resolution adjustment operator and the parallel attention mechanism are integrated to accurately focus on key anatomical structures; an end-to-end unified optimization framework from feature extraction to visualization output is constructed to realize automatic and fine segmentation and quantitative evaluation of uterine, ovarian and fallopian tube regions, and generate an analysis result image with abnormal region highlighting display and interactive operation, which significantly improves the accuracy and efficiency of gynecological image analysis and the intuitiveness of clinical diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0082] Figure 1 A basic flowchart of the gynecological image analysis system based on deep learning provided by an embodiment of the application is shown.

[0083] Figure 2 To improve the parallel attention structure diagram. DETAILED DESCRIPTION

[0084] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present application, not all embodiments.

[0085] Example 1, with reference to Figure 1 For an embodiment of the present application, a deep learning-based gynecological image analysis system is provided, including: a feature extraction module, a region division module, a feature analysis module, an abnormality evaluation module, and a visualization output module;

[0086] The feature extraction module is configured to perform multi-scale feature extraction on the original gynecological medical image through a multi-resolution convolutional neural network to obtain a multi-scale feature map;

[0087] The region division module is configured to divide the original gynecological medical image based on the multi-scale feature map to obtain a divided region;

[0088] The feature analysis module is configured to perform quantitative feature analysis on the divided region to obtain a feature description;

[0089] The abnormality evaluation module is configured to process the feature description to obtain an abnormality index;

[0090] The visualization output module is configured to visually mark the divided region based on the abnormality index to obtain an analysis result image.

[0091] Through the multi-resolution convolutional neural network and the cross-resolution fusion operation, efficient integration of multi-scale features is achieved, which enhances the perception ability of micro lesions and fuzzy boundaries while maintaining high-resolution details; the adaptive resolution adjustment operator and the parallel attention mechanism are integrated to accurately focus on key anatomical structures; an end-to-end unified optimization framework is constructed from feature extraction to visualization output, realizing automatic and refined segmentation and quantitative evaluation of the uterus, ovary and fallopian tube regions, and generating an analysis result image with abnormal region highlighting and interactive operation, significantly improving the accuracy and efficiency of gynecological image analysis and the intuitiveness of clinical diagnosis.

[0092] The processing logic for multi-scale feature extraction on the original gynecological medical image through the multi-resolution convolutional neural network includes:

[0093] The original gynecological medical image is input into the multi-resolution convolutional neural network, which is composed of multiple parallel branches;

[0094] Parallel branch is a parallel path for processing a specific resolution feature map independently in the multi-resolution convolutional neural network;

[0095] The first branch maintains the spatial resolution of the original gynecological medical image. In the multi-resolution convolutional neural network, a branch is added at each stage, and the spatial resolution of the added branch is half of the lowest resolution of the previous stage, and the number of channels is doubled.

[0096] At the end of each stage of the multi-resolution convolutional neural network, the output feature maps of all branches of the stage are subjected to a cross-resolution fusion operation.

[0097] After the cross-resolution fusion operation and the processing of the adaptive resolution adjustment operator in multiple stages of the multi-resolution convolutional neural network, the output feature maps of all branches of the last stage of the multi-resolution convolutional neural network are output as multi-scale feature maps to the region division module.

[0098] In a specific implementation, the feature extraction module realizes multi-scale feature extraction of gynecological medical images by constructing a multi-resolution convolutional neural network. A parallel multi-branch architecture is adopted, and the original gynecological medical image (with a size of 512x512 pixels) is input into the first branch and its spatial resolution is maintained, and the initial number of channels is set to 64. In the network, with each additional processing stage, a branch is added. The spatial resolution of the added branch is adjusted to half of the lowest resolution branch of the previous stage, and the number of channels is doubled. For example, a 256x256 resolution, 128-channel branch is added in the second stage; a 128x128 resolution, 256-channel branch is added in the third stage; and a 64x64 resolution, 512-channel branch is added in the fourth stage. The basic processing unit of each branch is composed of a 3x3 convolution layer, a batch normalization layer, and a ReLU activation function, which ensures that the network can expand the receptive field at each level while maintaining high-resolution detail perception, thereby significantly improving the detection capability of small lesions. At the end of each stage, a cross-resolution fusion operation is performed, which raises the low-resolution branch feature map to a high resolution through upsampling, while lowering the high-resolution branch feature map to the target resolution using downsampling. After 1x1 convolution for channel alignment, all branch features are aggregated using element-wise addition to form enhanced multi-scale feature representations. This operation improves the accuracy of organ boundary segmentation and the reliability of lesion classification. After fusion, an adaptive resolution adjustment operator is introduced, which dynamically optimizes the spatial structure of the feature map through a sampling point generator, calculates a dynamic range factor based on feature content, adjusts the original offset and generates a deformation sampling grid, and finally performs resampling using bilinear interpolation to enhance the feature response of lesion edges and small structures, effectively enhancing the signal-to-noise ratio of key areas and suppressing background interference. The last stage of all branch feature maps (512x512x64, 256x256x128, 128x128x256, 64x64x512) after multiple stage fusion and adaptive adjustment are output to the region division module as multi-scale feature maps, where the high-resolution feature map retains the endometrial texture details, and the low-resolution feature map encodes the global context of the organ, providing a complete feature basis covering micro to macro for subsequent region division. When the input image size changes, the spatial resolution of each branch will be scaled by the same principle, thereby ensuring the consistency of the multi-scale architecture.

[0099] The cross-resolution fusion operation specifically includes:

[0100] Each branch of the current stage is sequentially updated as the target branch;

[0101] Each branch of the current stage except the target branch is defined as a source branch, and the output feature map of each source branch is transformed to the same spatial resolution as the target branch through upsampling to obtain a transformed feature map with the same resolution as the target branch.

[0102] All transformed feature maps are aggregated with the output feature map of the target branch itself, and the output of the aggregation result is taken as the updated feature map of the target branch, and the updated feature map of the target branch is taken as the input of the target branch into the next stage;

[0103] After the cross-resolution fusion operation of the first stage to the third stage of the multi-resolution convolutional neural network, the output feature map of the corresponding branch is input to the adaptive resolution adjustment operator;

[0104] The adaptive resolution adjustment operator generates a set of sampling points for each input feature map through its internal sampling point generator, and performs a resampling operation on the input feature map using the set of sampling points to obtain an optimized feature map and update the feature representation of the corresponding branch.

[0105] In a specific implementation, the cross-scale transmission and enhancement of feature information is realized by constructing a multi-branch interaction mechanism, each branch of the current stage is determined as a target branch in turn, and taking a 512x512 high-resolution branch as an example, when updating the feature of the target branch, the 256x256, 128x128 and 64x64 branches of the same stage are defined as source branches. The output feature maps of each source branch are spatially transformed, among which the 256x256 branch is transformed by 2 times upsampling, the 128x128 branch is transformed by 4 times upsampling, and the 64x64 branch is transformed by 8 times upsampling, and all are transformed to the target resolution of 512x512, and then 1x1 convolution is performed for channel alignment, to ensure that all transformed feature maps and the target branch have the same dimension of 512x512x64. The transformed feature maps and the output feature map of the target branch are aggregated by element-wise addition, among which the high-resolution branch provides boundary detail information, the medium-resolution branch supplements regional texture features, and the low-resolution branch injects global semantic context, to form an enhanced feature representation with local precision and global perception, and the output of the aggregation result is taken as the output feature map of the updated target branch, which is taken as the input of the branch in the next stage, to ensure that each branch retains its own feature characteristics while fully absorbing complementary information of other scales. After the fusion operation in the first to third stages, the output feature maps of each branch are input into an adaptive resolution adjustment operator, which generates a deformation sampling grid for each input feature map through an internal sampling point generator, increases the sampling point density in key areas such as the endometrial boundary and the follicle wall, and performs resampling operation on the input feature map by bilinear interpolation to obtain a feature map with optimized spatial structure and update the feature representation of the corresponding branch. Under the synergistic effect of multi-scale feature complementation and adaptive optimization, the system significantly improves the global context understanding ability while maintaining the high-resolution detail perception, and through the cross-resolution fusion and adaptive optimization mechanism, the system integrates multi-scale features while maintaining the high-resolution detail perception ability, increases the detection rate and diagnostic accuracy of small lesions, and significantly enhances the accuracy and clinical practicability of gynecological image analysis.

[0106] The processing process of the adaptive resolution adjustment operator includes:

[0107] Performing global average pooling on the input feature map to obtain a channel feature vector;

[0108] Inputting the channel feature vector into the parameter prediction network to obtain a dynamic range factor;

[0109] Performing scalar multiplication operation on the dynamic range factor and the original offset generated by the convolution operation to obtain an adjusted offset;

[0110] Adding the adjusted offset and the original grid position by pixel reorganization technology to obtain a set of sampling points;

[0111] The original grid position refers to a normalized coordinate matrix with a same space size as the input feature map and a regular fixed size, which defines the initial space position of each pixel on the input feature map.

[0112] The input feature map is resampled by using a bilinear interpolation method based on a sampling point set to obtain an optimized feature map.

[0113] In the implementation, the adaptive resolution adjustment operator realizes the spatial structure optimization of the feature map by constructing a dynamic sampling mechanism. The input feature map is subjected to global average pooling processing, and the average value of all feature values of each channel in the spatial dimension is calculated, so that the feature map of 256x256x64 is compressed into a channel feature vector of 1x1x64. The channel feature vector is input into a parameter prediction network composed of a fully connected layer and a Sigmoid activation function, and a 64-dimensional dynamic range factor is output, the value range of which is controlled between 0 and 1. An original offset is generated by a 3x3 convolution operation, and a scalar multiplication operation is performed on the original offset and the dynamic range factor channel by channel to obtain an adjusted offset, wherein the high dynamic range factor corresponds to the lesion edge region, and the feature positioning accuracy is strengthened, and the high dynamic range factor is 0.7. The adjusted offset is added to the original grid position by a pixel recombination technology. The original grid position is a regular fixed normalized coordinate matrix (256x256) with a same space size as the input feature map, and the coordinate values are uniformly distributed in the interval [-1, 1], which defines the initial space position of each pixel on the input feature map. The input feature map is subjected to non-uniform resampling operation by using a bilinear interpolation method based on a sampling point set, and the adaptive improvement of the sampling point density (up to 2 times the standard density) is realized in the key areas such as the endometrial boundary and the follicle wall, and finally the feature map with optimized spatial structure is output. Through the dynamic sampling mechanism of the adaptive resolution adjustment operator, the representation ability of the lesion area in the gynecological image is significantly improved, and at the same time, the adaptive adjustment of the sampling density is realized through the parameter prediction network, which effectively enhances the perception accuracy of the model to the micro-lesions and complex boundaries on the premise of maintaining the calculation efficiency.

[0114] The parameter prediction network is composed of a fully connected layer and a Sigmoid activation function;

[0115] The fully connected layer is used to map the channel feature vector to the initial dynamic range factor;

[0116] The Sigmoid activation function is used to compress the value of the initial dynamic range factor to the interval of 0 to 1 to obtain the dynamic range factor.

[0117] In a specific implementation, the parameter prediction network is composed of a fully connected layer and a Sigmoid activation function, where the fully connected layer maps the 64-dimensional channel feature vector to a 64-dimensional initial dynamic range factor, and a bias-free linear transformation is used to ensure the simplicity of parameter prediction. The Sigmoid function is used to compress the value of the initial dynamic range factor to the interval [0, 1] to form the final dynamic range factor. During the training process, L2 regularization (weight decay 10 -4 ) is used to prevent overfitting. On a training set containing 15,000 ultrasound images, the parameter prediction network converges stably within 50 training cycles, and the dynamic range factor values are concentrated in the interval [0.2-0.8]. The high value area (>0.7) mainly corresponds to complex structure areas such as endometrial boundary and follicle wall, and the low value area (<0.3) corresponds to homogeneous tissue areas, achieving a balance between computational efficiency and representation ability. At the same time, in complex cases such as ovarian teratoma, the dynamic range factor of the hair dense area is automatically increased to 0.7-0.9, guiding the sampling points to gather in the key feature area, while the cystic area factor is maintained at 0.1-0.3, avoiding unnecessary computational overhead. The parameter prediction network effectively improves the feature enhancement efficiency of the adaptive resolution adjustment operator, while maintaining the lightweight model, and provides accurate spatial adjustment guidance for subsequent feature resampling.

[0118] The original gynecological medical image is divided into divided regions based on the multi-scale feature map, and the processing logic includes:

[0119] The multi-scale feature map is cross-scale feature fused based on the feature pyramid network to obtain a unified feature map;

[0120] The unified feature map is pixel-level semantic segmented based on the DeepLabv3+ architecture to obtain a semantic segmentation map;

[0121] It should be noted that the region division module constructs a multi-stage feature processing pipeline, and performs step-by-step fusion on four scale feature maps (512x512x64, 256x256x128, 128x128x256, 64x64x512) output by the multi-resolution convolutional neural network. Along the top-down path, the deep feature map of the 64x64 scale is obtained by 2 times upsampling to obtain the first scale fusion feature map, and the channel dimension of the shallow feature map of the corresponding level in the multi-scale feature map is unified by 1x1 convolution in the horizontal connection to obtain the second scale fusion feature map; the first scale fusion feature map and the second scale fusion feature map are added element by element to obtain the fusion feature of 128x128, 256x256 and 512x512 three levels, and the fusion feature is iteratively passed to the next level scale according to the path from 64x64 to 128x128, then to 256x256, and then to 512x512, realizing step-by-step feature fusion from top to bottom, and finally generating a unified feature map with a unified size of 256x256x256.

[0122] Based on the DeepLabv3+ architecture, the above unified feature map is processed for pixel-level segmentation, wherein the encoder extracts deep semantic features based on a pre-trained multi-resolution convolutional neural network, the segmentation decoder performs an atrous spatial pyramid pooling (ASPP) on the unified feature map, and further restores the spatial resolution by 3x3 convolution and 4 times upsampling, and predicts the class of each pixel point in the unified feature map, and outputs a semantic segmentation map containing three classes of uterus, ovary and fallopian tube.

[0123] Based on the semantic segmentation map, the connected component analysis algorithm is used to aggregate the pixel points belonging to the same class and spatially adjacent, to obtain an initial region set;

[0124] The morphological features and statistical features of the initial region set are calculated, and the initial region set is filtered and merged according to the preset filtering rule to obtain a division region;

[0125] Each division region includes a class label and corresponding geometric information, the class label includes uterus, ovary and fallopian tube, and the corresponding geometric information includes a pixel set of spatial position in the original gynecological medical image and a boundary coordinate sequence;

[0126] The division region includes a uterus region, an ovary region, and a fallopian tube region.

[0127] In practice, the semantic segmentation map was post-processed using a connected component analysis algorithm. An 8-neighbor connectivity criterion was used to aggregate spatially adjacent pixels of the same type into an initial region set, resulting in an average of 28.3 initial regions per image. Morphological features such as area (number of pixels), perimeter, roundness, and density were calculated for each initial region, and statistical features such as grayscale mean and texture entropy were extracted. Pre-defined filtering rules were applied: area threshold > 50 pixels, roundness threshold > 0.3, and density threshold > 0.5. Regions that did not meet these criteria were filtered out, and regions with spatial overlap > 70% were merged. This resulted in accurately segmented regions. Each segmented region was assigned a category label (uterus / ovary / fallopian tube) and corresponding geometric information, including the set of pixel coordinates constituting the region and the boundary coordinate sequence describing the region's contour. This effectively improved the Dice coefficient for uterine region segmentation. Particularly in cases of severe endometriosis with adhesions, the synergistic effect of multi-scale feature fusion and morphological filtering significantly reduced the missegmentation rate, providing a precise anatomical basis for subsequent feature analysis and anomaly assessment.

[0128] The processing logic for performing quantitative feature analysis on the divided regions to obtain feature descriptions includes:

[0129] An improved parallel attention structure is introduced when performing quantitative feature analysis on the segmented regions.

[0130] Improved parallel attention structures include channel attention units and spatial attention units;

[0131] The channel attention unit is used for multi-scale channel weighting based on multiple channel responses, and its calculation formula is as follows:

[0132] ;

[0133] ;

[0134] ;

[0135] ;

[0136] in, Indicates the first The weighted eigenvalues ​​of each channel, Indicates the first The global average pooling value for each channel. Indicates the first The global max pooling value for each channel. Indicates the first The pyramid pooling value of each channel, The weight parameters represent the global average pooling value. The weight parameters represent the global max pooling value. a weight parameter representing a pyramid pooling value, a height of a multi-scale feature map, a width of a multi-scale feature map, a horizontal coordinate of a spatial position, a vertical coordinate of a spatial position, a feature value of a th channel at a spatial position , a total number of channels, a multi-scale feature map of a th channel is subjected to a pyramid pooling operation, a multi-scale feature map of a th channel;

[0137] The calculation method of the multi-scale channel weight is:

[0138] ;

[0139] ;

[0140] ;

[0141] ;

[0142] wherein, a weight of a th channel, a weight of global average pooling, a weight of global maximum pooling, a weight of pyramid pooling, an activation function;

[0143] The spatial attention unit generates a spatial weight matrix based on multi-scale convolution;

[0144] The channel weight and the spatial weight are cross-weighted to obtain a dynamic weight distribution, and the calculation formula of the cross-weighting is:

[0145] ;

[0146] ;

[0147] wherein, a weight tensor, a spatial weight, a splicing operation, a dimension of a spliced weight tensor, a spliced weight tensor, a combination operation;​

[0148] The dynamic weight distribution is multiplied element-wise with the input feature map to obtain the enhanced feature representation, which is then used as the feature description.

[0149] It is important to emphasize that this application innovatively introduces an improved parallel attention structure through a feature analysis module to achieve quantitative feature enhancement and description of the segmented region. The channel attention unit performs multi-scale channel weighting based on multiple channel responses. First, it calculates the global average pooling value, global max pooling value, and four-level pyramid pooling value for each channel. Global average pooling obtains global contextual information by averaging across all positions in the feature map in the spatial dimension; global max pooling retains significant feature responses by selecting the maximum value in the feature map; and pyramid pooling captures spatial features of different granularities through multi-scale pooling windows (1×1, 2×2, 3×3, 6×6). The weighted fusion of the three pooling results uses learnable weight parameters, adaptively adjusting the contribution of each pooling method through end-to-end training. Finally, it generates multi-scale channel weights that can simultaneously perceive global statistical characteristics and local salient features. The calculation formula is as follows:

[0150] ;

[0151] ;

[0152] ;

[0153] ;

[0154] in, Indicates the first The weighted eigenvalues ​​of each channel, Indicates the first The global average pooling value for each channel. Indicates the first The global max pooling value for each channel. Indicates the first The pyramid pooling value of each channel, The weight parameters represent the global average pooling value. The weight parameters represent the global max pooling value. The weight parameters represent the pyramid pooling values. Indicates the height of the multi-scale feature map. This represents the width of the multi-scale feature map. The horizontal coordinate representing spatial location. The vertical coordinate represents the spatial location. Indicates the first The spatial location of each channel eigenvalues ​​at that location denotes the total number of channels, denotes the multi-scale feature map of the th channel, performing a pyramid pooling operation, denotes the multi-scale feature map of the th channel;

[0155] The calculation method of the multi-scale channel weight is:

[0156] ;

[0157] ;

[0158] ;

[0159] ;

[0160] wherein, denotes the weight of the th channel, denotes the weight of the global average pooling, denotes the weight of the global maximum pooling, denotes the weight of the pyramid pooling, denotes an activation function;

[0161] The spatial attention unit generates a spatial weight matrix based on multi-scale convolution, and uses 1×1, 3×3 and 5×5 convolution kernels of different sizes in parallel to process the input feature map, wherein the 1×1 convolution focuses on the integration of features across channels, the 3×3 convolution captures local spatial patterns, and the 5×5 convolution obtains more extensive context information. The convolution outputs of the three scales are spliced and fused by convolution to generate a spatial weight matrix that emphasizes key spatial positions, and the calculation formula is:

[0162] ;

[0163] ;

[0164] ;

[0165] wherein, denotes the input feature map, denotes the 1×1 convolution kernel weight, denotes the 3×3 convolution kernel weight, denotes the 5×5 convolution kernel weight;

[0166] The channel weight and the spatial weight are combined through a cross weighting mechanism, the channel weight is extended along the spatial dimension, and the spatial weight is spliced to form a multi-dimensional weight tensor, the interaction modeling of the channel dimension and the spatial dimension is realized through the combination operation, and finally the dynamic weight distribution is generated. The distribution is multiplied by the input feature map element by element to realize the synergistic enhancement of the key channel and the important spatial position, and finally the feature description is formed, and the cross weighting calculation formula is:

[0167] ;

[0168] ;

[0169] Wherein, represents the weight tensor, represents the spatial weight, represents the splicing operation, represents the dimension of the spliced weight tensor, represents the weight tensor obtained by splicing, represents the combination operation;

[0170] Through the synergistic effect of multi-scale channel weighting and spatial attention, the system effectively enhances the feature response of the lesion area while maintaining the integrity of the features, providing a feature representation with higher information density and stronger discriminability for subsequent anomaly evaluation.

[0171] Based on the feature description, an abnormality index is generated, and the processing logic includes:

[0172] Local feature quantities are extracted from the feature description, including regional internal texture contrast, regional internal gradient amplitude statistics, and feature difference between sub-regions;

[0173] The quantized features in the feature description are statistically processed to obtain a feature value set;

[0174] Discrete measure parameters are calculated for the feature value set, and the discrete measure parameters and the local feature quantities are calculated through linear weighting method to obtain intermediate evaluation quantities;

[0175] The intermediate evaluation quantities are calculated by normalization to obtain initial abnormal quantities;

[0176] Based on the feature set of the divided region, a regional level statistic is established, and the regional level statistic and the initial abnormal quantity are associated to obtain a regional weight quantity;

[0177] The regional weight quantity and the initial abnormal quantity are linearly combined to obtain an abnormality index;

[0178] The abnormality index includes regional anomaly score, abnormal heat value, regional risk level and abnormal distribution matrix.

[0179] In specific implementation, three key feature quantities are extracted from the feature description: the texture contrast within a region is calculated using Local Binary Pattern (LBP) variance, and micro-texture changes are captured through 3×3 neighborhood analysis; the gradient magnitude statistics within a region are calculated based on the Sobel operator to calculate the gradients in the x and y directions, and extract their mean, variance, and kurtosis. The gradient mean reflects the clarity of the organizational boundary, the variance characterizes the complexity of the internal structure, and the kurtosis describes the concentration of the gradient distribution; the feature difference between sub-regions is achieved by uniformly dividing each region into 4×4 sub-blocks, calculating the cosine similarity matrix between the depth features of each sub-block, and then quantifying the consistency within the region based on the statistical properties of this matrix.

[0180] Distribution analysis was performed on the 128-dimensional quantified features in the feature description, calculating their mean, standard deviation, skewness, and kurtosis to form a feature value set. Three discreteness metrics—variance, information entropy, and Gini coefficient—were then calculated for this set. Variance characterizes the degree of dispersion of the feature distribution, information entropy quantifies the complexity of the features, and the Gini coefficient reflects the balance of the feature distribution. The discreteness metrics were integrated with local feature quantities using a linear weighting method. The weighting coefficients were determined through grid search as follows: texture contrast 0.25, gradient statistics 0.35, feature dissimilarity 0.20, and discreteness parameter 0.20, ultimately yielding a comprehensive intermediate evaluation metric.

[0181] The Min-Max normalization method was used to map intermediate evaluation values ​​to the [0,1] interval to obtain initial anomaly values. Regional statistics were established based on the feature set of the divided regions, including region area weight (log(1+Area)), spatial centrality weight (calculated based on the distance of the region from the image center), and category importance weight (uterus 0.4, ovary 0.35, fallopian tube 0.25). These statistics were correlated with the initial anomaly values ​​to obtain region weights. A linear combination of the region weights and the initial anomaly values ​​was performed to output a four-dimensional anomaly index, including a region anomaly score (0-1 continuous values), anomaly heat value (normalized to grayscale values ​​in the [0,255] range), region risk level (divided into three levels based on thresholds: low [0,0.3], medium [0.3,0.7], and high [0.7,1.0]), and anomaly distribution matrix (a 256×256 two-dimensional probability map).

[0182] Through collaborative analysis and hierarchical quantification of multi-dimensional features, the system achieves accurate assessment of abnormalities in obstetric and gynecological imaging, providing reliable quantitative basis for clinical diagnosis and meeting the timeliness requirements of real-time diagnosis.

[0183] Visual labeling of segmented regions based on anomaly indicators yields the analysis result image. The processing logic includes:

[0184] Get each partitioned region;

[0185] According to the abnormal index of each divided region, a display color is assigned to it through a predefined color mapping rule, and in a blank matrix with the same size as the original gynecological medical image, color filling is performed at the corresponding position of the image according to the geometric information, to obtain a region marking layer;

[0186] According to the geometric information of each divided region, the corresponding abnormal index value is filled into all pixel positions of the region in the image, to obtain a spatial abnormal distribution map;

[0187] The spatial abnormal distribution map is subjected to spatial interpolation processing to obtain an abnormal heat map;

[0188] The region marking layer and the abnormal heat map layer are superimposed on the original gynecological medical image in a semi-transparent manner, to obtain an analysis result image.

[0189] In a specific implementation, the geometric information (including a pixel set and a boundary coordinate sequence) of each divided region is obtained, and a display color is assigned to the region abnormal score through a predefined Jet color spectrum mapping rule: a low-risk region (score 0-0.3) is mapped to a blue system, a medium-risk region (0.3-0.7) is mapped to a yellow-green system, and a high-risk region (0.7-1.0) is mapped to a red system. In a blank matrix with the same size as the original 512*512 medical image, accurate color filling is performed according to the region geometric information to generate a region marking layer, and a 2-pixel-wide deep red boundary is additionally added to the high-risk region to enhance visual warning.

[0190] The abnormal index value corresponding to each divided region is filled into all pixel positions of the region to generate an initial spatial abnormal distribution map. A bicubic interpolation algorithm is used to smooth the spatial abnormal distribution map, and the mutation at the region boundary is eliminated through weighted calculation of 16 adjacent pixels to generate a continuously and gradually changing abnormal heat map, which significantly improves the naturalness of the visual effect.

[0191] Alpha blending technology is used to superimpose the region marking layer and the abnormal heat map layer on the original image in a semi-transparent manner, wherein the region marking layer is set to have a transparency of 0.4, the abnormal heat map layer is set to have a transparency of 0.6, and visual optimization is realized through the following blending formula:

[0192] ;

[0193] wherein, represents an output pixel, represents a pixel value of the original gynecological medical image, represents a pixel value of the region marking layer, represents a pixel value of the abnormal heat map, and a transparency control parameter representing the overall transparency of the region marker layer and the abnormal heat map, a region marker layer transparency, , an abnormal heat map transparency, and ;

[0194] The analysis result image effectively highlights the potential lesion area while maintaining the clear and identifiable original anatomical structure. This visualization method significantly improves the training efficiency of the resident physician, and the analysis time of complex cases is significantly shortened.

[0195] The visualization output module further includes an abnormal region highlighting process, and the processing logic thereof includes:

[0196] Taking the region abnormal score in the abnormal index as the judgment basis, the region abnormal score is compared with the preset high abnormal threshold value;

[0197] If the region abnormal score of the divided region is greater than or equal to the preset high abnormal threshold value, it is determined as a high abnormal region;

[0198] If the region abnormal score of the divided region is less than the preset high abnormal threshold value, it is determined as a normal region;

[0199] Based on the analysis result image, the high abnormal region is displayed with special marking different from the ordinary region for enhanced display;

[0200] The special marking includes dynamic flashing effect, thickened boundary contour line and superimposed preset warning symbol;

[0201] The analysis result image is interactively adjusted, and the interactive adjustment includes zooming, rotating and superimposing different levels of abnormal information.

[0202] In a specific implementation, taking the region abnormal score in the abnormal index as the judgment basis, the score of each divided region is compared with the preset high abnormal threshold value 0.75, which is determined based on ROC curve analysis. At this threshold value, the sensitivity of the system to malignant lesions is greater than 90% and the specificity is greater than 85%. When the region abnormal score is greater than or equal to the threshold value, the system determines it as a high abnormal region, otherwise it is marked as a normal region.

[0203] Based on the analysis result image, a triple special marking is performed on the high abnormal area: a dynamic flashing effect is realized by CSS animation, red-yellow color alternately displayed, frequency set to 2Hz, visual residence time controlled to 500ms; a boundary contour line is thickened from standard 1 pixel to 3 pixels, drawn with a deep red (255, 0, 0) solid line, ensuring that it is still clear and visible under a complex tissue background; a warning symbol is superimposed on the region centroid position with a 16x16 pixel exclamation mark icon, the icon design adopts a high-contrast scheme of red background and white text, and the symbol transparency is set to 0.8 to avoid completely blocking the underlying anatomical structure.

[0204] A multi-dimensional operation function is provided to support real-time interactive adjustment of the analysis result image: a zoom operation is realized by a bicubic interpolation algorithm to ensure that the image details are not distorted in the zoomed state; a rotation function supports free rotation from 0° to 360°, with a minimum adjustment unit of 5°, facilitating physicians to observe the lesion morphology from multiple angles; an information superimposition function allows selective display of different levels of abnormal information, including separate display of the region marker layer, the abnormal heat map, or the superimposed mode of the two, and physicians can quickly switch the display combination through a check box interface. This highlighting display scheme significantly shortens the first fixation time of physicians on high-risk lesions, and gradually reduces the diagnosis decision time.

[0205] Those skilled in the art will appreciate that embodiments of the present application can be provided as methods, systems or computer program products. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (media) having computer-usable program code embodied thereon. The storage media can be any available memory or storage device, including both volatile and non-volatile memory, as well as other types of storage. For example, the storage media can include, but is not limited to, RAM, EEPROM, flash memory, a hard disk, or other forms of storage. These computer program instructions can be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device that implements the flowchart block or blocks. The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operations to be performed on the computer or other programmable data processing apparatus to produce a computer implemented process such that the instructions executed by the computer or other programmable data processing apparatus provide operations for implementing the flowchart block or blocks.Figure 1 one or more processes and / or functions specified in a block Figure 1 one or more blocks.

[0206] It should be noted that the above-mentioned embodiments are only used to illustrate the technical solutions of the present application, not to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalent replaced without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A deep learning-based obstetrics and gynecology image analysis system, characterized by, The method comprises the following steps: The feature extraction module is used to perform multi-scale feature extraction on the original gynecological medical image through a multi-resolution convolutional neural network to obtain a multi-scale feature map; The region division module is used to divide the original gynecological medical image based on the multi-scale feature map to obtain a division region; The feature analysis module is used to perform quantitative feature analysis on the division region to obtain a feature description; The abnormality evaluation module is used to process the feature description to obtain an abnormality index; The visual output module is used to visually mark the division region based on the abnormality index to obtain an analysis result image. The processing logic for performing multi-scale feature extraction on the original gynecological medical image through a multi-resolution convolutional neural network to obtain a multi-scale feature map comprises:

2. The deep learning based obstetrics and gynecology image analysis system of claim 1, wherein, The original gynecological medical image is input into the multi-resolution convolutional neural network, which is composed of multiple parallel branches; The parallel branch is a parallel path in the multi-resolution convolutional neural network for independently processing a specific resolution feature map; At the end of each stage of the multi-resolution convolutional neural network, the output feature maps of all branches of the stage are subjected to a cross-resolution fusion operation; After the cross-resolution fusion operation and the processing of the adaptive resolution adjustment operator in multiple stages of the multi-resolution convolutional neural network, the output feature maps of all branches of the last stage of the multi-resolution convolutional neural network are obtained as the multi-scale feature map and output to the region division module. The cross-resolution fusion operation specifically comprises: Each branch of the current stage is sequentially updated as a target branch; 3. The deep learning based obstetrics and gynecology image analysis system of claim 2, wherein, Each branch of the current stage except the target branch is defined as a source branch, and the output feature map of each source branch is transformed to the same spatial resolution as the target branch through up-sampling to obtain a transformed feature map aligned with the resolution of the target branch; The aggregated result of the output of the aggregated operation is used as the updated feature map of the target branch, and the updated feature map of the target branch is used as the input of the next stage of the target branch; After the cross-resolution fusion operation in the first to third stages of the multi-resolution convolutional neural network, the output feature maps of the corresponding branches are input to the adaptive resolution adjustment operator; The adaptive resolution adjustment operator generates a set of sampling points for each input feature map through its internal sampling point generator, and performs resampling operation on the input feature map using the set of sampling points to obtain an optimized feature map and update the feature representation of the corresponding branch. The processing process of the adaptive resolution adjustment operator comprises: Performing global average pooling on the input feature map to obtain a channel feature vector; 4. The deep learning based obstetrics and gynecology image analysis system of claim 3, wherein: Inputting the channel feature vector into a parameter prediction network to obtain a dynamic range factor; ​ ​ The dynamic range factor is multiplied by the original offset generated by the convolution operation to obtain an adjusted offset; The adjusted offset is added to the original grid position through a pixel reorganization technique to obtain a set of sampling points; The original grid position refers to a normalized coordinate matrix with the same spatial size as the input feature map and a fixed rule, which defines the initial spatial position of each pixel on the input feature map; The input feature map is resampled using a bilinear interpolation method based on the set of sampling points to obtain an optimized feature map.

5. The deep learning based obstetric and gynecologic image analysis system of claim 4, wherein: The parameter prediction network is composed of a fully connected layer and a Sigmoid activation function; The fully connected layer is used to map the channel feature vector to an initial dynamic range factor; The Sigmoid activation function is used to compress the value of the initial dynamic range factor to the interval of 0 to 1 to obtain the dynamic range factor.

6. The deep learning based obstetrics and gynecology image analysis system of claim 1, wherein, The original gynecological medical image is divided into divided regions based on the multi-scale feature map, and the processing logic includes: The multi-scale feature map is cross-scale feature fused based on the feature pyramid network to obtain a unified feature map; The unified feature map is subjected to pixel-level semantic segmentation based on the DeepLabv3+ architecture to obtain a semantic segmentation map; Based on the semantic segmentation map, the connected component analysis algorithm is used to aggregate the pixels belonging to the same category and spatially adjacent to obtain an initial region set; The initial region set is calculated for morphological features and statistical features, and the initial region set is filtered and merged according to the preset filtering rule to obtain the divided region; Each divided region includes a category label and corresponding geometric information, the category label includes the uterus, ovary and fallopian tube, and the corresponding geometric information includes a pixel set and a boundary coordinate sequence of the spatial position in the original gynecological medical image; The divided region includes the uterus region, the ovary region and the fallopian tube region.

7. The deep learning based obstetrics and gynecology image analysis system of claim 1, wherein, The processing logic for obtaining the feature description through quantitative feature analysis of the divided region includes: An improved parallel attention structure is introduced when performing quantitative feature analysis on the divided region; The improved parallel attention structure includes a channel attention unit and a spatial attention unit; The channel attention unit is used for multi-scale channel weighting based on multiple channel responses, and the calculation formula is: ; ； ； ； wherein, represents a weighted feature value of the th channel, represents a global average pooling value of the th channel, represents a global max pooling value of the th channel, represents a pyramid pooling value of the th channel, represents a weight parameter of the global average pooling value, represents a weight parameter of the global max pooling value, represents a weight parameter of the pyramid pooling value, represents a height of the multi-scale feature map, represents a width of the multi-scale feature map, represents an abscissa of a spatial position, represents an ordinate of a spatial position, represents a feature value of the th channel at the spatial position , represents a total number of channels, represents that a pyramid pooling operation is performed on the multi-scale feature map of the th channel, , represents a multi-scale feature map of the th channel; The calculation method of the multi-scale channel weight is: ； ； ； ； wherein, represents the weight of the th channel, represents the weight of the global average pooling, represents the weight of the global max pooling, represents the weight of the pyramid pooling, represents an activation function; The spatial attention unit generates a spatial weight matrix based on multi-scale convolution; The channel weight and the spatial weight are cross-weighted to obtain a dynamic weight distribution, and the calculation formula of the cross-weighting is: ； ； wherein, denotes a weight tensor, denotes a spatial weight, denotes a concatenation operation, denotes a dimension of the concatenated weight tensor, denotes a concatenated weight tensor, denotes a combination operation; The dynamic weight distribution is multiplied element by element with the input feature map to obtain an enhanced feature representation, which is taken as the feature description.

8. The deep learning based obstetric and gynecologic image analysis system of claim 1, wherein, The processing logic for generating an abnormality indicator based on the feature description includes: Local feature quantities are extracted from the feature description, including the internal texture contrast of the region, the statistical quantity of the internal gradient amplitude of the region, and the feature difference between sub-regions; The distribution statistics of the quantitative features in the feature description are obtained to obtain a feature value set; The discrete measure parameter of the feature value set is calculated, and the discrete measure parameter and the local feature quantity are calculated by linear weighting method to obtain an intermediate evaluation quantity; The initial abnormality quantity is calculated by normalizing the intermediate evaluation quantity; The region-level statistics are established based on the feature set of the divided region, and the region weight quantity is calculated by associating the region-level statistics and the initial abnormality quantity; The abnormality index is obtained by linearly combining the region weight quantity and the initial abnormality quantity; The abnormality index includes region abnormality score, abnormal heat value, region risk level and abnormal distribution matrix.

9. The deep learning based obstetrics and gynecology image analysis system of claim 1, wherein, The divided region is visually marked based on the abnormality index to obtain an analysis result image, and the processing logic includes: Obtaining each divided region; According to the abnormality index of each divided region, a display color is assigned to it through a pre-defined color mapping rule, and a region marking layer is obtained by filling the color in the corresponding position of the image according to the geometric information of the region in a blank matrix with the same size as the original gynecological medical image. According to the geometric information of each divided region, the corresponding abnormality index value is filled into all pixel positions of the region in the image to obtain a spatial abnormality distribution map. The spatial abnormality distribution map is subjected to spatial interpolation processing to obtain an abnormal heat map. The region marking layer and the abnormal heat map layer are superimposed on the original gynecological medical image in a semi-transparent manner to obtain an analysis result image.

10. The deep learning based obstetric and gynecologic image analysis system of claim 9, wherein, The visualization output module further includes abnormal region highlight display processing, and the processing logic includes: The region abnormality score in the abnormality index is used as a judgment basis to compare the region abnormality score with a pre-set high abnormality threshold value. If the region abnormality score of the divided region is greater than or equal to the pre-set high abnormality threshold value, it is determined as a high abnormality region. If the region abnormality score of the divided region is less than the pre-set high abnormality threshold value, it is determined as a normal region. Based on the analysis result image, the high abnormality region is displayed in a special mark different from the ordinary region for enhanced display. The special mark includes a dynamic flashing effect, a thickened boundary contour line and a superimposed pre-set warning symbol. The analysis result image is subjected to interactive adjustment, and the interactive adjustment includes zooming, rotating and superimposing different levels of abnormal information.

Images