Mammary gland molybdenum target image intelligent analysis system and method based on multi-scale deep learning

Abstract

The invention relates to the technical field of medical image intelligent analysis systems, in particular to a mammary gland molybdenum target image intelligent analysis system and method based on multi-scale deep learning, and the system comprises an image obtaining module which is used for obtaining a mammary gland molybdenum target image; the calcification detection module is connected with the image acquisition module, and is used for carrying out intelligent detection on calcification points in the mammary gland molybdenum target image based on the multi-dimensional feature matrix and the adaptive feature filter, and generating confidence scores and dynamic segmentation thresholds of the calcification points; the calcification segmentation module is connected with the calcification detection module and is used for segmenting the calcification point region based on the dynamic segmentation threshold value of the calcification detection module; the calcification quantification module is connected with the calcification segmentation module and is used for extracting quantitative features of the calcification points based on the segmentation result of the calcification segmentation module, and by introducing feature extraction and feature fusion of different scales, the sensitivity to tiny calcification points is remarkably improved; and meanwhile, the problem that a traditional single-scale model cannot accurately identify calcification characteristics of different scales is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image intelligent analysis systems, and in particular to a breast mammography image intelligent analysis system and method based on multi-scale deep learning. Background Art

[0002] Currently, breast cancer screening and early diagnosis increasingly rely on intelligent analysis of mammography images. Calcification points, as early lesion features of breast cancer, have high diagnostic value. However, due to the diversity of the size, density and shape of calcification points in mammography images, traditional detection methods based on single-scale feature extraction are easily affected by background noise, morphological differences and densely distributed areas in complex breast tissues. These methods mostly use single-scale convolutional neural networks (CNNs) or fixed threshold segmentation, which cannot accurately identify complex calcification features and have technical problems such as high missed detection rate, high false detection rate and low segmentation accuracy.

[0003] In the existing single-scale convolutional network model, the detection of calcification points mainly relies on the feature extraction of a single scale or direction. This method is often powerless for tiny or densely distributed calcification points. At the same time, the traditional method lacks a flexible feature weighting and segmentation threshold adjustment mechanism, which easily misidentifies non-calcified areas in breast mammography images as calcification points, thereby increasing the false positive rate. Fixed segmentation thresholds are usually unable to adapt to calcification points of different densities and sizes. Especially when segmenting calcification points in dense areas, the blurred boundaries between adjacent calcification points lead to inaccurate segmentation results, which are difficult to meet clinical needs.

[0004] In addition, traditional methods lack adaptability in feature extraction and weighting, and fail to assign dynamic weights to calcification points according to the importance of features at different scales and directions. Existing feature extraction methods usually do not have sufficient suppression capabilities for background noise, resulting in a high false detection rate when detecting tiny calcification points. Due to the lack of adaptive feature filters, the system performs poorly in dealing with complex backgrounds and high-noise environments, limiting the sensitivity and accuracy of calcification point detection. In addition, the computational efficiency of existing methods is relatively low, making it difficult to quickly process a large number of breast mammography images, which is not conducive to actual clinical applications. Summary of the invention

[0005] In response to the above-mentioned problems of the prior art, the present invention provides a breast mammography image intelligent analysis system and method based on multi-scale deep learning, aiming to achieve high-precision calcification point detection through the coordinated cooperation of multiple modules, combined with multi-scale feature extraction and dynamic segmentation threshold adjustment.

[0006] The present invention proposes a breast mammography image intelligent analysis system based on multi-scale deep learning, comprising:

[0007] An image acquisition module, used for acquiring breast mammography images;

[0008] A calcification detection module, connected to the image acquisition module, is used to intelligently detect calcification points in the mammary gland mammography image based on a multidimensional feature matrix and an adaptive feature filter, and to generate a confidence score and a dynamic segmentation threshold for the calcification points;

[0009] A calcification segmentation module, connected to the calcification detection module, for segmenting the calcification point area based on the dynamic segmentation threshold of the calcification detection module;

[0010] The calcification quantification module is connected to the calcification segmentation module and is used to extract quantitative features of calcification points based on the segmentation results of the calcification segmentation module.

[0011] Preferably, the image acquisition module performs image preprocessing on the breast mammography image, and the image preprocessing includes denoising and image enhancement to improve the clarity of the image and the detection accuracy of calcification points.

[0012] Preferably, the calcification detection module comprises:

[0013] The steps of constructing a multidimensional feature matrix are used to extract the features of calcification points in mammography images from different directions and scales, and generate a feature matrix M d，s , where d is the direction dimension, s is the scale dimension, (x, y) is the pixel coordinate, and the element M in the feature matrix d，s (x, y) represents the feature value of the pixel point (x, y) in direction d and scale s;

[0014] Use adaptive feature filter to filter the feature matrix M d，s Perform feature screening to obtain the denoised feature matrix F d，s , the filter removes low-intensity background noise based on a set feature threshold T;

[0015] Calculate the adaptive feature weight W d，s , used to allocate the weight of calcification point features in multi-scale features, where α is the direction adjustment parameter and β is the scale adjustment parameter;

[0016] Generating a calcification point confidence score C(x, y) based on the adaptive feature weights, for quantifying the accuracy of the detection results;

[0017] A dynamic segmentation threshold λ(x, y) is generated based on the confidence score and used for the segmentation operation of the calcification point.

[0018] Preferably, in the step of constructing a multidimensional feature matrix, the feature matrix is ​​generated in the following manner:

[0019]

[0020] Where f(x, y) is the pixel intensity value in the mammography image, g d，s (x, y) is a two-dimensional feature basis function in direction d and scale s, which is used to extract the corresponding calcification point feature information.

[0021] Preferably, the adaptive feature filter is used for the feature matrix M d，s Filter to obtain the denoised feature matrix F d，s ,in

[0022] F ds (x, y) = M d，s (x, y)·δ(x, y),

[0023] And δ(x, y) is the filter function, when M d，s When (x, y)>T, δ(x, y)=1, otherwise δ(x, y)=0, where T represents a characteristic threshold value for filtering low-intensity background noise.

[0024] Preferably, the adaptive feature weight W d，s Based on the denoised feature matrix F d，s Calculation is performed to determine the weight distribution of calcification points in multi-scale features, where

[0025]

[0026] The α represents a direction sensitivity adjustment parameter, and the β represents a scale sensitivity adjustment parameter, which are used to adjust the influence of different directions and scales on the calcification point.

[0027] Preferably, the calcification point confidence score C(x, y) is based on the adaptive feature weight W d，s and the calcification point detection result S(x, y) are generated to quantify the accuracy of the calcification point detection result, where

[0028]

[0029] The γ is a confidence smoothing coefficient, S(x, y)=1 indicates that a calcification point is detected, and S(x, y)=0 indicates that no calcification point is detected.

[0030] Preferably, the dynamic segmentation threshold λ(x, y) is dynamically adjusted based on the calcification point confidence score C(x, y) for accurate segmentation of the calcification point area, wherein

[0031] λ(xy)=θ·C avg (1+C(x, y)),

[0032] The C avgis the average confidence of the breast mammography image, and θ is a threshold adjustment factor, which is used to adjust the segmentation threshold to adapt to calcification points of different densities and shapes.

[0033] The intelligent analysis method of mammary gland mammography images based on multi-scale deep learning based on the system comprises the following steps:

[0034] Generate a multidimensional feature matrix M based on the mammography image d，s ; Use an adaptive feature filter to filter the feature matrix to obtain a denoised feature matrix, and calculate the adaptive feature weight W of the feature matrix d，s ; Generate a calcification point confidence score C(x, y) based on the feature weight; Generate a dynamic segmentation threshold λ(x, y) based on the confidence score; Segment the calcification point area based on the dynamic segmentation threshold; Extract quantitative features of the calcification point.

[0035] Preferably, the mammographic target image is subjected to image preprocessing after being obtained, and the image preprocessing includes denoising and enhancement to improve the quality of image input and enhance the detection of calcification point features.

[0036] The beneficial technical effects brought by the technical solution of the present invention are:

[0037] The system of the present invention includes an image acquisition module, a calcification detection module, a calcification segmentation module and a calcification quantification module. The modules are connected consistently through data flow and information flow to form a highly coordinated closed-loop system structure, thereby ensuring the overall stability and detection accuracy of the system.

[0038] The multi-scale feature pyramid method proposed in the present invention significantly improves the sensitivity to tiny calcification points in the calcification detection module by introducing feature extraction and feature fusion at different scales, while overcoming the problem that the traditional single-scale model cannot accurately identify calcification features at different scales. On the basis of calcification point feature extraction, the present invention adopts an adaptive feature filter to remove background noise and retain high-intensity calcification point information, further suppressing false positives. In particular, in terms of feature weighting, the system generates adaptive feature weights based on the importance of features in different directions and scales, ensuring that the system can accurately weight the features of calcification points at multiple scales and angles, greatly enhancing the robustness of calcification point detection.

[0039] To improve the accuracy of segmentation, the calcification detection module of the present invention dynamically generates a segmentation threshold based on the confidence score of the calcification point to adapt to the characteristics of calcification points of different densities and shapes. The dynamic segmentation threshold solves the problem that the fixed threshold is prone to blurred boundaries in the detection of dense calcification areas, ensuring that the system can more accurately locate the boundaries of the calcification area. At the same time, the present invention designs an effective data processing flow in the collaboration of multiple modules and the closed loop of information flow, and forms a strict logical closure between the modules in information transmission and input and output alignment, so that the entire system can efficiently process image data while ensuring detection accuracy.

[0040] From a macroscopic perspective, the system of the present invention, under the synergistic effect of multiple modules, uses multi-scale feature extraction, adaptive feature weighting and dynamic threshold segmentation to achieve high-precision detection, segmentation and quantitative analysis of calcification points. The close coupling of multiple modules and the smooth connection of data flow significantly improve the detection performance of the system and achieve higher accuracy in complex backgrounds. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is the overall logic block diagram of the system of the present invention.

[0042] Figure 2 This is the core algorithm logic block diagram of the present invention.

[0043] Figure 3 Figure 1. Segmentation of microcalcifications and breast masses in a left breast mammogram of a patient with invasive ductal carcinoma. (a) The craniocaudal (CC) view shows focally clustered microcalcifications (indicated by thin arrows) and an irregular round mass (indicated by thick arrows). (b) The suspicious mass is automatically delineated within the red curve. (c) The segmented microcalcifications detected in (b) are used to characterize the features. DETAILED DESCRIPTION

[0044] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects thereof are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0045] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0046] See also Figure 1-3The present invention provides a breast mammography image intelligent analysis system and method based on multi-scale deep learning. The system of the present invention includes an image acquisition module 1, a calcification detection module 2, a calcification segmentation module 3 and a calcification quantification module 4. The image acquisition module 1 is used to acquire the breast mammography image and transmit it to the calcification detection module 2. The calcification detection module 2 detects the calcification points in the breast mammography image based on a multidimensional feature matrix and an adaptive feature filter, and generates a confidence score of the calcification point and a dynamic segmentation threshold. After receiving the segmentation threshold output by the calcification detection module 2, the calcification segmentation module 3 adjusts the segmentation accuracy of the calcification point area to achieve accurate segmentation of the calcification point. Finally, the calcification quantification module 4 extracts the quantitative features of the calcification point, such as the area and quantity, based on the segmented calcification area, to provide a strong basis for subsequent medical diagnosis.

[0047] Preferably, the image acquisition module 1 can perform appropriate preprocessing operations based on the specific quality of the mammographic target image. For example, the image can be enhanced by an adaptive histogram equalization method to improve the contrast, making the calcification points more clearly visible in the image, thereby being more helpful for subsequent detection by the calcification detection module 2. Such a modular structure ensures that the input and output between the modules are consistent, so that the image information can be transmitted efficiently and accurately, forming an integrated intelligent analysis system.

[0048] Preferably, the image acquisition module 1 may also include the function of denoising and enhancing the mammary gland molybdenum target image. The denoising process may be performed based on Gaussian filtering or adaptive filtering to smooth out high-frequency noise in the image, thereby improving the clarity of the image. For the enhancement process, an adaptive contrast enhancement algorithm may be used to enhance the significance of the calcification point, making the calcification area clearer in the image, thereby enabling more accurate detection and analysis of the calcification point features. Such a preprocessing method not only effectively improves the input quality of the image, but also ensures the detection performance of the system in a complex background environment.

[0049] In the embodiment of the present invention, the calcification detection module 2 of the present invention first constructs a multidimensional feature matrix for extracting the features of calcification points in the mammography image from different directions and scales to generate a feature matrix M d，s , where d is the direction dimension, s is the scale dimension, (x, y) is the pixel coordinate, and the element M in the feature matrix d，s (x, y) represents the feature value of the pixel point (x, y) in direction d and scale s;

[0050] Use adaptive feature filter to filter the feature matrix M d，s Perform feature screening to obtain the denoised feature matrix F d，s , the filter removes low-intensity background noise based on a set feature threshold T;

[0051] Calculate the adaptive feature weight W d，s , used to allocate the weight of calcification point features in multi-scale features, where α is the direction adjustment parameter and β is the scale adjustment parameter;

[0052] Generating a calcification point confidence score C(x, y) based on the adaptive feature weights, for quantifying the accuracy of the detection results;

[0053] Based on the confidence score, a dynamic segmentation threshold λ(x, y) is generated for the segmentation of calcification points. The multi-scale feature matrix designed in this way effectively covers calcification features of different sizes and directions, so that the system can accurately identify calcification points in various situations, especially showing excellent detection effects for micro-calcification features.

[0054] Preferably, in the step of constructing a multidimensional feature matrix, the feature matrix is ​​generated in the following manner:

[0055]

[0056] Where f(x, y) is the pixel intensity value in the mammography image, g d , s(x, y) is a two-dimensional feature basis function in direction d and scale s, which is used to extract the corresponding calcification point feature information.

[0057] Next, the adaptive feature filter removes background noise mainly based on the feature value threshold T, leaving the high-intensity calcification feature area.

[0058] Preferably, the present invention will automatically adjust the eigenvalue threshold T according to the changes in image brightness and contrast during the generation of the feature matrix. Generally, the threshold T can be 1.5 to 2.0 times the average grayscale of the image to ensure that low-intensity noise is effectively removed. The filtered feature matrix F d，s It is expressed as

[0059] F ds (x, y) = M d，s (x, y)·δ(x, y),

[0060] And δ(x, y) is the filter function, when M d，s When (x, y)>T, δ(x, y)=1, otherwise δ(x, y)=0, where T represents the feature threshold, which is used to filter low-intensity background noise. This design can significantly reduce false detections in calcification point detection and improve the signal-to-noise ratio of calcification features.

[0061] In an embodiment of the present invention, the adaptive feature weight W d，s The calculation method is based on the filtered feature matrix F d，s, and is used to distribute weights between features of different scales and directions. The specific formula is:

[0062]

[0063] The α represents the direction sensitivity adjustment parameter, and the β represents the scale sensitivity adjustment parameter, which is used to adjust the influence of different directions and scales on calcification points, and is usually set to 1.0 and 1.5 respectively, so as to achieve a reasonable weight distribution among the features. Through the adaptive calculation of feature weights, the system can automatically focus on the most representative calcification features, and further improve the recognition effect of calcification points while suppressing background interference.

[0064] In an embodiment of the present invention, based on the adaptive feature weight W d，s , the calcification detection module 2 generates a confidence score C(x, y) of the calcification point, and the formula is as follows:

[0065]

[0066] The γ is a confidence smoothing coefficient, S(x, y) = 1 indicates that a calcification point is detected, and S(x, y) = 0 indicates that a calcification point is not detected, wherein γ is usually set between 0.8 and 1.0 to smooth the scoring results. The confidence score effectively quantifies the detection accuracy of calcification points by combining feature weights and detection results.

[0067] In an embodiment of the present invention, the system generates a dynamic segmentation threshold λ(x, y) based on the confidence score C(x, y) to accurately segment the calcification area. Specifically, the segmentation threshold formula is:

[0068] λ(xy)=θ·C avg (1+C(x,y))

[0069] Where θ is the threshold adjustment factor, which is usually set in the range of 0.5 to 2.0, C avg Represents the average confidence value of the image. This dynamic threshold can automatically adapt to the morphology and density changes of calcification points, making the segmentation result more accurate, especially suitable for dense calcification areas.

[0070] In the embodiment of the present invention, after receiving the segmentation threshold of the calcification detection module 2, the calcification segmentation module 3 segments the calcification point area according to the dynamic segmentation threshold λ(x, y). The segmentation process can adjust the segmentation accuracy of each area according to the confidence of the calcification point to ensure accurate coverage of the calcification area boundary, thereby avoiding false detection and missed detection.

[0071] In an embodiment of the present invention, a method for intelligent analysis of mammary gland molybdenum target images based on multi-scale deep learning is also disclosed, comprising the following steps:

[0072] Generate a multidimensional feature matrix M based on the mammography image d，s ; Use an adaptive feature filter to filter the feature matrix to obtain a denoised feature matrix, and calculate the adaptive feature weight W of the feature matrix d，s ; Generate a calcification point confidence score C(x, y) based on the feature weight; Generate a dynamic segmentation threshold λ(x, y) based on the confidence score; Segment the calcification point area based on the dynamic segmentation threshold; Extract the quantitative features of the calcification point. The gradual implementation of the above method enables the system to further extract the quantitative features of the calcification point after the calcification point area is segmented with high precision, thereby realizing intelligent breast image analysis.

[0073] Preferably, the mammographic target image is first preprocessed by denoising and contrast enhancement after acquisition. The contrast enhancement operation may preferably use an adaptive contrast enhancement technique to significantly highlight the calcification features, thereby ensuring the detection accuracy of the system in a complex background environment.

[0074] In order to verify the superiority of the intelligent analysis system and method of mammography images based on multi-scale deep learning of the present invention, public professional data sets such as Digital Data base for Screening Mam mography (DDSM) are used for testing. The DDSM data set contains a large number of detailed annotated mammography images. It is a standard test set for breast calcification detection and analysis and is widely used in the research of computer-aided detection and diagnosis of breast lesions. This comparative test selected 200 mammography images of different types and features, and tested key indicators such as the accuracy of calcification detection, segmentation accuracy, accuracy of quantitative feature extraction, and system computational efficiency to verify the actual application effect of the method of the present invention in complex mammography image analysis.

[0075] In the embodiment of the present invention, calcification point feature extraction based on multi-scale feature pyramid, adaptive adjustment of filters and dynamic segmentation threshold mechanism are used to process the DDSM dataset. For each mammography image, accurate identification and quantitative analysis of calcification points are achieved by adjusting and optimizing the following key parameters:

[0076] 1. The scale range of the feature pyramid: multi-layer pooling is performed at scales of 1x1, 2x2, 3x3, 6x6, and 7x7.

[0077] 2. Adaptive feature filter threshold T: set to 1.5 times the average grayscale of the image to filter out background noise.

[0078] 3. Feature weight parameter W d，s:Adaptively adjust according to the different scales of calcification points to balance the contribution of each scale feature to calcification point detection.

[0079] 4. Confidence score smoothing coefficient γ: set to 0.9 to smooth the confidence score of calcification point detection.

[0080] 5. Dynamic segmentation threshold adjustment coefficient θ: set to 1.5 to ensure accurate segmentation of the calcification area.

[0081] This embodiment applies the above parameters and processes to the DDSM dataset to obtain high-precision calcification point detection results.

[0082] The comparison adopts the traditional calcification point detection method based on a single-scale convolutional neural network (CNN) and is tested in the DDSM dataset. The traditional CNN method is easy to miss small calcification points in the image and fails to make adaptive adjustments in feature weight allocation, noise filtering, etc. In addition, the comparison does not use a dynamic segmentation threshold, but a fixed threshold, which leads to a high false detection rate and inaccurate segmentation in dense calcification areas.

[0083] The test standards and methods are as follows:

[0084] The test mainly examines the following key indicators:

[0085] Detection accuracy (Accuracy): The ratio of the number of correctly detected calcification points to the total number of calcification points.

[0086] Segmentation Precision: The overlap ratio between the segmented calcification area and the actual area.

[0087] False Positive Rate (FPR): The false positive rate of non-calcification points.

[0088] Computational efficiency (Processing Time per Image, PTI): The processing time per image, in seconds (s).

[0089] Among them, the detection accuracy and segmentation precision are calculated by the standard overlap rate of calcification point detection: the overlap between the detected calcification point area and the annotated area exceeds 70% and is considered a successful detection. The false positive rate is obtained through statistical analysis of the false positive rate. Each detection indicator is performed under the same hardware conditions to ensure the fairness of the comparison results.

[0090] The test results are shown in the following table:

[0091] index Embodiments of the present invention Comparative Example Detection accuracy (%) 94.8 82.3 Segmentation accuracy (%) 91.5 75.4 False positive rate (%) 4.3 12.8 Computational efficiency (PTI, s) 0.9 1.7

[0092] It can be seen from the test results that the present invention is significantly better than the comparative example in terms of detection accuracy, segmentation accuracy and false positive rate. First, in terms of detection accuracy, the present invention effectively covers the features of calcification points of different scales and directions through the application of multi-scale feature pyramids, so that the detection accuracy of calcification points reaches 94.8%, which is significantly improved compared with the 82.3% of the comparative example. Especially in terms of segmentation accuracy, the dynamic segmentation threshold adjustment mechanism of the present invention effectively improves the accuracy of calcification area segmentation, so that the segmentation accuracy reaches 91.5%, which is 16.1 percentage points higher than the comparative example, proving the superiority of adaptive dynamic segmentation. In addition, the significant decrease in the false positive rate (only 4.3%) shows that the adaptive feature filter of the present invention plays a significant role in reducing the impact of background noise. Finally, due to the optimization of multi-scale structure and weight adjustment, the computational efficiency of the present invention is also better than that of traditional methods, and the processing time of each image is shortened to 0.9 seconds, which is faster than the comparative example method. This embodiment verifies the superiority of the present invention in breast mammography image analysis.

[0093] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. Intelligent analysis system of mammary gland mammography images based on multi-scale deep learning, characterized by: include: An image acquisition module, used for acquiring breast mammography images; A calcification detection module, connected to the image acquisition module, is used to intelligently detect calcification points in the mammary gland mammography image based on a multidimensional feature matrix and an adaptive feature filter, and to generate a confidence score and a dynamic segmentation threshold for the calcification points; A calcification segmentation module, connected to the calcification detection module, for segmenting the calcification point area based on the dynamic segmentation threshold of the calcification detection module; The calcification quantification module is connected to the calcification segmentation module and is used to extract quantitative features of calcification points based on the segmentation results of the calcification segmentation module.

2. The intelligent analysis system for mammographic target images based on multi-scale deep learning according to claim 1, characterized in that: The image acquisition module performs image preprocessing on the breast molybdenum target image, and the image preprocessing includes denoising and image enhancement to improve the clarity of the image and the detection accuracy of the calcification point.

3. The intelligent analysis system for mammographic target images based on multi-scale deep learning according to claim 1, characterized in that: The calcification detection module comprises: The steps of constructing a multidimensional feature matrix are used to extract the features of calcification points in mammography images from different directions and scales, and generate a feature matrix M d,s , where d is the direction dimension, s is the scale dimension, (x, y) is the pixel coordinate, and the element M in the feature matrix d,s (x,y) represents the feature value of the pixel point (x,y) in direction d and scale s; Use adaptive feature filter to filter the feature matrix M d,s Perform feature screening to obtain the denoised feature matrix F d,s , the filter removes low-intensity background noise based on a set feature threshold T; Calculate the adaptive feature weight W d,s , used to allocate the weight of calcification point features in multi-scale features, where α is the direction adjustment parameter and β is the scale adjustment parameter; Generating a calcification point confidence score C(x,y) based on the adaptive feature weights to quantify the accuracy of the detection results; A dynamic segmentation threshold λ(x,y) is generated based on the confidence score for segmentation of calcification points.

4. The breast mammography image intelligent analysis system based on multi-scale deep learning according to claim 3, characterized in that: In the step of constructing a multidimensional feature matrix, the feature matrix is ​​generated in the following manner: Where f(x,y) is the pixel intensity value in the mammography image, g d,s (x, y) is a two-dimensional feature basis function in direction d and scale s, which is used to extract the corresponding calcification point feature information.

5. The breast platinum target image intelligent analysis system based on multi-scale deep learning according to claim 3 is characterized in that: The adaptive feature filter is used to filter the feature matrix M d,s Filter to obtain the denoised feature matrix F d,s ,in F ds (x,y)=M d,s (x,y)·δ(x,y), And δ(x,y) is the filter function, when M d,s When (x, y)>T, δ(x, y)=1, otherwise δ(x, y)=0, where T represents a characteristic threshold used to filter low-intensity background noise.

6. The intelligent analysis system for mammographic target images based on multi-scale deep learning according to claim 5, characterized in that: The adaptive feature weight W d,s Based on the denoised feature matrix F d,s Calculation is performed to determine the weight distribution of calcification points in multi-scale features, where The α represents a direction sensitivity adjustment parameter, and the β represents a scale sensitivity adjustment parameter, which are used to adjust the influence of different directions and scales on the calcification point.

7. The intelligent analysis system for mammographic target images based on multi-scale deep learning according to claim 3, characterized in that: The calcification point confidence score C(x, y) is based on the adaptive feature weight W d,s and the calcification point detection result S(x,y) are generated to quantify the accuracy of the calcification point detection result, where The γ is a confidence smoothing coefficient, S(x, y)=1 indicates that a calcification point is detected, and S(x, y)=0 indicates that no calcification point is detected.

8. The breast platinum target image intelligent analysis system based on multi-scale deep learning according to claim 7, characterized in that: The dynamic segmentation threshold λ(x,y) is dynamically adjusted based on the calcification point confidence score C(x,y) for accurate segmentation of the calcification point area, where λ(xy)=θ·C avg ·(1+C(x,y)), The C avg is the average confidence of the breast mammography image, and θ is a threshold adjustment factor, which is used to adjust the segmentation threshold to adapt to calcification points of different densities and shapes.

9. A method for intelligent analysis of mammographic images based on multi-scale deep learning based on the system according to any one of claims 1 to 8, characterized in that: The following steps are involved: Generate a multidimensional feature matrix M based on the mammography image d,s ; Use an adaptive feature filter to filter the feature matrix to obtain a denoised feature matrix, and calculate the adaptive feature weight W of the feature matrix d,s ; Generate a calcification point confidence score C(x,y) based on the feature weight; Generate a dynamic segmentation threshold λ(x,y) based on the confidence score; Segment the calcification point area based on the dynamic segmentation threshold; Extract quantitative features of the calcification point.

10. The method for intelligent analysis of mammary gland molybdenum target images based on multi-scale deep learning according to claim 9, characterized in that: The mammographic target image is preprocessed after being obtained, and the image preprocessing includes denoising and enhancement to improve the quality of image input and enhance the detection of calcification point features.

Images