Deep learning based micro-calcification detection method for mammogram

Abstract

The present application relates to the technical field of calcification point segmentation, and particularly relates to a breast molybdenum target image micro-calcification point identification and detection method based on deep learning. The present application performs image segmentation on a breast molybdenum target image based on each segmentation threshold, and determines a suspected micro-calcification point region under each segmentation threshold. Based on the morphological features, light and dark features and texture features of the suspected micro-calcification point region, a feature vector is determined, and a trained neural network is used to determine the confidence score of the suspected micro-calcification point region. The confidence scores of each pixel point in the suspected micro-calcification point region under different segmentation thresholds are corrected and fused to determine a final score and determine a final micro-calcification point region. Through the cooperative verification and fusion of multiple thresholds, the present application eliminates local tissue interference when the global threshold is determined, and improves the detection sensitivity of micro-calcification point identification and the accuracy of boundary extraction.

Description

Technical Field

[0001] This invention relates to the field of calcification segmentation technology, specifically to a method for identifying and detecting microcalcifications in mammograms based on deep learning. Background Technology

[0002] Breast microcalcifications are a typical manifestation of calcium salt deposition in breast tissue on X-ray imaging. Their core clinical value lies in the fact that they may be the only or most important imaging feature of early breast cancer, especially ductal carcinoma in situ before it has formed a visible mass. Certain microcalcifications with specific morphologies and distribution patterns are highly correlated with malignant lesions. Therefore, accurate and early identification of microcalcifications in mammography is crucial for improving the early diagnosis rate of breast cancer and enhancing patient prognosis.

[0003] Currently, global thresholding is a common technique for identifying calcifications in images. However, microcalcifications are small in size and have low contrast, often appearing blurry in images and easily confused with structures such as blood vessel sections and tissue overlap artifacts. Furthermore, the background of breast tissue, such as glands and fat, has complex textures and grayscale fluctuations, severely interfering with the accuracy of segmentation. Setting a high threshold can reduce false positives, but it will lead to the missed detection of a large number of tiny, low-contrast real calcifications. Setting a low threshold can improve the detection rate, but it will introduce a large amount of noise and false positives from normal tissue structures, significantly increasing the review burden on clinicians, thus resulting in low detection accuracy for microcalcifications. Summary of the Invention

[0004] To address the technical problem of low detection accuracy of microcalcifications due to the inability of global threshold segmentation to adapt to the gray-level heterogeneity and morphological diversity of breast background, this invention aims to provide a deep learning-based method for identifying and detecting microcalcifications in mammograms. The specific technical solution adopted is as follows: This invention proposes a deep learning-based method for identifying and detecting microcalcifications in mammograms, the method comprising: Obtain mammograms of the test subjects; Several segmentation thresholds were determined; based on each segmentation threshold, the mammogram was segmented to identify the suspected microcalcification areas under each segmentation threshold. Based on the morphological, light and dark, and texture features of the suspected microcalcification point region, a feature vector is determined; based on the feature vector, a trained neural network is used to determine the confidence score of the suspected microcalcification point region. The confidence scores of each pixel in the suspected microcalcification region under different segmentation thresholds are corrected and fused using all segmentation thresholds to determine the final score of each pixel in the mammogram; based on the final score, the final microcalcification region is determined.

[0005] Further, determining the feature vector includes: Calculate the standard deviation of the distances from the centroid of the suspected microcalcification region to all edge pixels, denoted as the shape dispersion; count the total number of pixels in the suspected microcalcification region, denoted as the region area; the morphological features include the shape dispersion and the region area. The brightness contrast is determined based on the grayscale difference between the suspected microcalcification area and its surrounding area; the brightness contrast is included in the light and dark features. The gradient values ​​of edge pixels in the suspected microcalcification region are obtained, and the average gradient values ​​of all edge pixels are calculated to obtain the edge saliency; the information entropy of the grayscale values ​​of pixels in the suspected microcalcification region is obtained and denoted as the texture saliency; the texture features include the edge saliency and the texture saliency. A feature vector is formed by the shape dispersion, area, brightness contrast, edge saliency, and texture saliency of the suspected microcalcification point region.

[0006] Further, determining the brightness contrast includes: For each edge pixel in the suspected calcification point region, draw a ray passing through the edge pixel with the centroid of the suspected microcalcification point region as the endpoint; extend the ray by a preset distance to obtain the analysis boundary point; Connect all edge pixels on the edge of the suspected microcalcification area to the corresponding analysis boundary points in sequence to form a closed boundary line; the area between the edge of the suspected microcalcification area and the boundary line is denoted as the background ring. The difference between the average gray value of pixels in the suspected microcalcification area and the average gray value of pixels in the background ring is calculated. The difference is used as the numerator, and the ratio obtained by dividing the average gray value of pixels in the background ring by a preset positive number is used as the denominator. This ratio is denoted as the brightness contrast ratio.

[0007] Further, determining the final score for each pixel in the mammogram includes: Select a target threshold for each pixel in the mammogram, where the pixel is located within the suspected microcalcification area below the target threshold. The ratio of each target threshold for each pixel to the sum of all target thresholds is used as the threshold weight for that target threshold. Using the threshold weights of each target threshold for each pixel, the confidence scores of the pixel in the suspected microcalcification area under the corresponding target threshold are weighted and summed to obtain the final score of the pixel.

[0008] Furthermore, determining the final microcalcification point region includes: The connected region formed by pixels in the mammogram whose final score is greater than a preset first score threshold is denoted as the calcification core region. For each calcified core region, a morphological expansion operation is performed on the calcified core region to obtain an expanded region; the edge pixels of the calcified core region are used as the initial growth points, and the region is grown under the condition that the final score is greater than a preset second score threshold to obtain the final microcalcified point region.

[0009] Furthermore, the method for obtaining the segmentation threshold includes: The optimal threshold is obtained by using the maximum inter-class variance method for the gray values ​​of pixels in the mammogram. The optimal threshold is used as the minimum calcification threshold, and the maximum gray value of the pixels in the image is selected as the maximum calcification threshold. Within the grayscale range formed by the minimum calcification threshold and the maximum calcification threshold, scanning is performed with a preset step size to obtain different segmentation thresholds.

[0010] Furthermore, the suspected microcalcification region under each segmentation threshold is a connected region composed of pixels in the mammogram whose gray value is greater than or equal to the segmentation threshold.

[0011] Furthermore, the preset first score threshold is greater than the preset second score threshold.

[0012] Furthermore, the trained neural network is a fully connected neural network; its training process includes: obtaining feature vectors containing real microcalcifications and false positive samples, where false positive samples include malignant calcifications and vascular artifacts; and using the cross-entropy function as the loss function.

[0013] Furthermore, the preset step size is less than the difference between the maximum calcification threshold and the minimum calcification threshold.

[0014] The present invention has the following beneficial effects: Firstly, the neural network comprehensively analyzes the morphological, brightness, and texture features of suspected microcalcification areas, effectively learning and distinguishing the structural differences between genuine microcalcifications and false positive targets such as vascular artifacts, tissue overlap, and noise. Through intelligent screening based on multi-dimensional features, it can eliminate false signals introduced by low segmentation thresholds, effectively solving the problem of false detection caused by low segmentation thresholds and improving the reliability of detection results.

[0015] Secondly, a multi-threshold segmentation strategy is adopted to dynamically adapt the local contrast between different calcifications and surrounding tissues, ensuring comprehensive coverage of various calcifications and effectively improving the detection sensitivity of small and blurred calcifications. The segmentation threshold is the most direct evidence of calcifications derived from pixel grayscale features. The confidence scores of the suspected microcalcification regions of each pixel under different segmentation thresholds are corrected and fused using each segmentation threshold. Through multi-scale collaborative verification, local tissue interference during global threshold determination is eliminated, so that the final score can more accurately reflect the true physical morphology of microcalcifications. Thus, more complete and accurate microcalcifications can be determined based on the final score, solving the problem of low detection accuracy caused by grayscale heterogeneity and calcification morphology diversity in breast images. Attached Figure Description

[0016] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 The flowchart illustrates a method for identifying and detecting microcalcifications in mammograms based on deep learning, as provided in one embodiment of the present invention. Figure 2 A flowchart illustrating a method for obtaining a feature vector according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a computer device for identifying and detecting microcalcifications in mammograms based on deep learning, as provided in one embodiment of the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a deep learning-based method for identifying and detecting microcalcifications in mammography images proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0019] Unless otherwise defined, 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 pertains.

[0020] The following description, in conjunction with the accompanying drawings, details a specific scheme for the identification and detection method of microcalcifications in mammograms based on deep learning, provided by this invention.

[0021] Example 1: This invention proposes a deep learning-based method for identifying and detecting microcalcifications in mammograms. Please refer to [link to relevant documentation]. Figure 1 The diagram illustrates a flowchart of a deep learning-based method for identifying and detecting microcalcifications in mammograms, according to an embodiment of the present invention. The method includes: Step S1: Obtain mammograms of the test subject.

[0022] Medical staff guided the test subject to remain still and hold their breath for several seconds to avoid image blurring caused by respiratory movements. Subsequently, an initial mammogram of the breast area was acquired using a fully digital mammogram. Finally, to reduce interference from fat, glandular tissue, and other tissues while preserving high-frequency edge information of minute calcifications and enhancing the contrast between calcifications and surrounding breast tissue, the initial mammogram was denoised and enhanced to obtain a final mammogram. The grayscale value of each pixel in the mammogram corresponds to the attenuation coefficient of the tissue at that point for X-rays.

[0023] It should be noted that in the embodiments of the present invention, non-local mean filtering is used for noise reduction and histogram equalization is used for image enhancement. The specific methods are not described here, but are all technical means well known to those skilled in the art.

[0024] Step S2: Determine several segmentation thresholds; perform image segmentation on the mammogram based on each segmentation threshold to determine the suspected microcalcification area under each segmentation threshold.

[0025] Accurate segmentation of microcalcifications is limited by the gray-level heterogeneity of the background tissue and the instability of the gray-level of the calcifications themselves, resulting in poor performance of globally fixed threshold methods. High thresholds may miss tiny or low-contrast calcifications, while low thresholds are prone to falsely detecting noise and normal structures. Furthermore, global thresholds struggle to handle the gray-level gradients within calcifications, easily leading to fragmentation or adhesion to neighboring points, compromising the accuracy of the segmentation morphology. Therefore, it is necessary to determine multiple segmentation thresholds.

[0026] Mammogram images are segmented based on various segmentation thresholds to identify suspected microcalcification regions at each threshold. This multi-threshold segmentation strategy dynamically adapts to the local contrast between different calcifications and surrounding tissues, ensuring comprehensive coverage of various types of calcifications and effectively improving the detection sensitivity for small and blurred calcifications.

[0027] Step S3: Determine the feature vector based on the morphological, light and dark, and texture features of the suspected microcalcification point region; based on the feature vector, use the trained neural network to determine the confidence score of the suspected microcalcification point region.

[0028] Microcalcifications typically appear as small, regularly shaped circular or oval spots. Due to their high density, calcifications are always brighter than the surrounding background in X-ray images; this brightness directly reflects their physical nature. True microcalcifications are relatively homogeneous internally, while false positive areas such as areas with tissue overlap, calcification hyperplasia, and noise exhibit complex internal structures and textures. In summary, morphological, brightness, and texture features constitute a comprehensive descriptive system from macro to micro and from external to internal perspectives. By fusing these features into a single feature vector, neural networks can learn the complex combinations and trade-offs between these features, thereby achieving a more accurate and robust confidence assessment of suspected microcalcification areas, obtaining a confidence score, and effectively reducing false positives and false negatives.

[0029] Step S4: Use all segmentation thresholds to correct and fuse the confidence scores of the suspected microcalcification regions of each pixel under different segmentation thresholds to determine the final score of each pixel in the mammogram; based on the final score, determine the final microcalcification region.

[0030] Due to the complex background of breast tissue and the diverse morphologies of microcalcifications, a single fixed threshold is insufficient to adapt to calcification features under different contrast levels, easily leading to oversegmentation or undersegmentation. The confidence score only derives from the neural network's overall judgment of suspected microcalcification regions, losing information about the grayscale distribution within these regions. In contrast, the segmentation threshold, derived from pixel grayscale features, is the most direct evidence of calcifications; a larger segmentation threshold corresponds to more obvious calcification features. This paper corrects and fuses the confidence scores of each pixel's suspected microcalcification regions under different segmentation thresholds using all segmentation thresholds. Through multi-scale collaborative verification, local tissue interference during global threshold determination is eliminated, enabling the final score to more accurately reflect the true physical morphology of microcalcifications and improving the detection sensitivity and boundary extraction accuracy of microcalcification identification.

[0031] Preferably, in some possible implementations of the embodiments of the present invention, the method for obtaining the segmentation threshold includes: using the maximum inter-class variance method to obtain the optimal threshold for the gray values ​​of pixels in the mammogram, using the optimal threshold as the minimum calcification threshold, and selecting the maximum gray value of pixels in the image as the maximum calcification threshold; scanning within the gray range formed by the minimum calcification threshold and the maximum calcification threshold with a preset step size to obtain different segmentation thresholds. The maximum inter-class variance method is a well-known technique and will not be described in detail here.

[0032] It should be noted that the optimal threshold can separate meaningful tissues such as glands and potential calcifications from background elements such as air and extremely dark areas. Calcifications appear as white bright spots, and their grayscale values ​​are necessarily higher than the optimal threshold. Therefore, the optimal threshold can be used as the lower limit of the threshold search range. The maximum calcification threshold can cover all possible brightness levels in the image, without missing any extremely bright calcifications. Using a preset step size for scanning to determine the segmentation threshold ensures accuracy and improves computational efficiency.

[0033] In one implementation of this invention, the preset step size is set to 15, which ensures that there are at least two segmentation thresholds. The implementer can set the step size according to the specific circumstances.

[0034] Preferably, in some possible implementations of the embodiments of the present invention, the method for obtaining the suspected microcalcification region includes: forming a connected region composed of pixels in a mammogram whose grayscale value is greater than or equal to each segmentation threshold, thereby constituting the suspected microcalcification region under that segmentation threshold. It should be noted that breast calcifications are calcium salt deposits in tissue, which appear as high-brightness points in mammograms; the suspected microcalcification region needs to be composed of pixels whose grayscale value is greater than or equal to the segmentation threshold.

[0035] Preferably, in some possible implementations of the embodiments of the present invention, the method for obtaining the feature vector is described in [reference needed]. Figure 2 The diagram illustrates a flowchart of a feature vector acquisition method according to an embodiment of the present invention, the method comprising: Step S210: Calculate the standard deviation of the distance from the centroid of the suspected microcalcification point region to all edge pixels, denoted as shape dispersion; count the total number of pixels in the suspected microcalcification point region, denoted as region area; morphological features include shape dispersion and region area.

[0036] It should be noted that the smaller the shape dispersion, the smaller the difference in boundary distances of suspected microcalcification areas in all directions, and the closer the shape is to a regular circle. Microcalcifications typically appear as small, regularly shaped, round or oval, high-brightness spots; malignant calcifications appear as larger, irregularly shaped, high-brightness areas, while vascular artifacts typically appear as linear structures. In this regard, area can distinguish between microcalcifications and malignant calcifications, while shape dispersion can distinguish between microcalcifications and vascular artifacts.

[0037] Step S220: Determine the brightness contrast based on the grayscale difference between the suspected microcalcification area and its surrounding area; the brightness and darkness features include brightness contrast.

[0038] Preferably, in some possible implementations of the embodiments of the present invention, the method for obtaining brightness contrast includes: for each edge pixel in the suspected calcification point region, drawing a ray passing through the edge pixel with the centroid of the suspected microcalcification point region as the endpoint; extending the ray by a preset distance to obtain the analysis boundary point; sequentially connecting all edge pixels on the edge of the suspected microcalcification point region to the corresponding analysis boundary point to form a closed boundary line; denoting the region between the edge of the suspected microcalcification point region and the boundary line as the background ring; calculating the difference between the average gray value of the pixels in the suspected microcalcification point region and the average gray value of the pixels in the background ring, using the difference as the numerator, and the ratio obtained by summing the average gray value of the pixels in the background ring and a preset positive number as the denominator, and denoting it as the brightness contrast. The preset positive number is used to prevent the denominator from being zero, thus making the fraction meaningless; in this embodiment, it is set to 0.01. It should be noted that calcification points are usually brighter than the surrounding tissue, and true calcification points need sufficient contrast to be distinguished from the background. By emitting rays from the centroid of suspected microcalcification areas and extending them to construct closed background rings, a precise, adaptive, and shape-fitting segmentation of the local background of calcification points is achieved, effectively avoiding the feature distortion problem caused by the inclusion of irrelevant tissue in the background area, a problem inherent in the traditional rectangular window method. Brightness contrast measures the prominence of suspected microcalcification areas relative to their neighboring areas. The greater the brightness contrast, the more likely a suspected microcalcification is a true microcalcification.

[0039] In one implementation of this invention, a preset distance is used to limit the pixel width of the background ring. This distance is determined by the implementer based on the spatial resolution of the mammogram and the prior empirical value of the microcalcification size. It is preferably set to 5 pixels to ensure that the background ring contains a sufficient number of pixel samples and to strictly limit the background sampling range to the adjacent spatial domain of the microcalcification, thus preventing the background ring from extending too far and crossing into the edges of nearby blood vessels, glands, or other heterogeneous tissue structures.

[0040] It should be noted that if the analysis boundary point obtained by extending along the ray direction exceeds the image boundary, the intersection of the ray and the image boundary will be taken as the analysis boundary point.

[0041] Step S230: Obtain the gradient values ​​of the edge pixels in the suspected microcalcification area, and average the gradient values ​​of all edge pixels to obtain the edge saliency; obtain the information entropy of the gray values ​​of the pixels in the suspected microcalcification area, denoted as the texture saliency; texture features include edge saliency and texture saliency.

[0042] It should be noted that genuine microcalcifications are relatively uniform internally and have sharp, clear edges, indicating low texture complexity and significant changes in boundary pixels, meaning a high boundary gradient. Regions with blurred boundaries or poor integration with the background have lower boundary gradient magnitudes and more complex textures. Greater edge saliency and lower texture saliency indicate a clearer and more uniform boundary in the suspected microcalcification area, making it more likely to be a genuine microcalcification.

[0043] In one implementation of this invention, the Canny operator is used to obtain the gradient values ​​of edge pixels.

[0044] Step S240: A feature vector is constructed from the shape dispersion, area, brightness contrast, edge saliency, and texture saliency of the suspected microcalcification point region.

[0045] It is important to note that elements with the same subscript in the feature vectors of different suspected microcalcification areas belong to the same data type. Shape dispersion, region area, brightness contrast, edge saliency, and texture saliency are denoted as analysis indicators. Before constructing the feature vector from these indicators, it is necessary to standardize each analysis indicator for a single suspected microcalcification area using the min-max normalization method, based on all analysis indicators for all suspected microcalcification areas. The feature vector is then constructed from all analysis indicators for the same suspected microcalcification area.

[0046] In this embodiment of the invention, a fully connected neural network is used to determine the confidence score of a suspected microcalcification region. The input of the fully connected neural network is the feature vector of the suspected microcalcification region, and the output is the confidence score of the suspected microcalcification region. The confidence score ranges from 0 to 1, representing the probability that the suspected microcalcification region is identified as a real microcalcification. The higher the score, the greater the likelihood that it is a real microcalcification.

[0047] The training process of a fully connected neural network is as follows: A pre-labeled training dataset is constructed, containing real microcalcification samples and various false positive samples (such as malignant calcifications and vascular artifacts). Each sample is labeled with a classification label: label 1 represents a true positive calcification, and label 0 represents a false positive calcification. The feature vector of the training samples is obtained using the same method as that for the suspected microcalcifications. The feature vector of the training samples is input into the fully connected neural network, which performs forward propagation calculations through multiple fully connected layers, outputting the predicted probability. The cross-entropy function is used as the loss function to calculate the error between the predicted probability and the classification label, and backpropagation is used to update the network parameters, resulting in a trained fully connected neural network. In use, the feature vector of the suspected microcalcification region to be processed is input into the trained fully connected neural network to output the corresponding confidence score.

[0048] Preferably, in some possible implementations of the embodiments of the present invention, the method for obtaining the final score includes: selecting a target threshold for each pixel in a mammogram, wherein the pixel is located within a suspected microcalcification region below the target threshold; using the ratio of each target threshold of each pixel to the sum of all target thresholds as the threshold weight of the target threshold; and using the threshold weights of each target threshold of each pixel to perform a weighted summation of the confidence scores of the pixel in the suspected microcalcification region below the corresponding target threshold to obtain the final score of the pixel.

[0049] In one implementation of this invention, the final score is expressed by the formula: In the formula, F is the final score for each pixel; K is the total number of target thresholds for each pixel. The k-th target threshold for each pixel; The nth target threshold for each pixel; The confidence score of the suspected microcalcification region for each pixel at the kth target threshold; Threshold weight for the k-th target threshold for each pixel.

[0050] It should be noted that the target threshold for a pixel reflects its eligibility to be considered a candidate for calcification at different gray levels. Since a larger target threshold implies that the pixel possesses more significant high-brightness features, the suspected microcalcification area is less affected by background breast tissue, resulting in a more reliable confidence score. Therefore, the threshold weight is determined by the target threshold size, allowing high-brightness features to dominate the fusion analysis. Through this weighting mechanism, a high target threshold provides a high-confidence determination of the core microcalcification area, while a low target threshold supplements the morphological features of the blurred edges of the microcalcification area, thereby achieving effective fusion of multi-scale segmentation information and ensuring that the final score accurately reflects the probability that a pixel belongs to a true microcalcification.

[0051] It should be noted that when a pixel does not belong to any suspected microcalcification area below the target threshold, the final score of that pixel is directly set to a constant of 0.

[0052] Preferably, in some possible implementations of the embodiments of the present invention, the method for obtaining the final microcalcification point region includes: defining the connected regions formed by pixels in the mammogram whose final score is greater than a preset first score threshold as calcification core regions; performing a morphological dilation operation on each calcification core region to obtain a dilated region; using the edge pixels of the calcification core region as initial growth points, and using a final score greater than a preset second score threshold as the growth condition, performing region growth to obtain the final microcalcification point region. The region growth algorithm is a well-known technique to those skilled in the art and will not be described in detail here.

[0053] It should be noted that the preset first score threshold provides high-quality seed points for subsequent region growth, reducing false positive diffusion. The calcified core region represents the most typical and reliable part of the microcalcifications. Since microcalcifications often appear in clusters, but may break due to threshold segmentation, the dilation operation can connect spatially adjacent but broken calcifications, ensuring that the growth process can cover the complete morphological range of the microcalcifications. The region growth process in this embodiment is as follows: any edge pixel of the calcified core region is used as a growth point. Pixels with a final score greater than the preset second score threshold in the eight-neighborhood of the growth point are used as new growth points for region growth. Growth stops when the final scores of all pixels in the eight-neighborhood of all new growth points are less than or equal to the preset second score threshold, resulting in the final microcalcified region. The preset second score threshold allows for the inclusion of pixels with lower final scores that may still be calcified, capturing surrounding blurry or low-contrast parts to obtain the final microcalcified region. Therefore, the preset first score threshold is greater than the preset second score threshold. This embodiment employs a dual-threshold strategy, proceeding from the inside out and from the core area to the complete region, ensuring both the specificity of the detection and the integrity of the microcalcification point boundaries.

[0054] In one implementation of this invention, a preset first score threshold is used to determine the initial seed point for region growth. Setting a higher value can strictly filter out false positive high-score noise caused by vascular artifacts or tissue overlap, ensuring extremely high reliability of calcification core region localization. The preset first score threshold is set to a range of 0.7 to 0.9, preferably 0.7. A preset second score threshold is used to control the termination condition of region growth. Setting a lower value allows the growth algorithm to adaptively spread outward, encompassing real blurred pixels at the edge of microcalcification points that have low confidence due to partial volume effect or low contrast but still have morphological continuity, thereby completely restoring the true boundary of microcalcification points. The preset second score threshold is set to a range of 0.2 to 0.5, preferably 0.3. Implementers can choose values ​​within the above range based on actual verification data or the distribution histogram of neural network output scores.

[0055] This invention is now complete.

[0056] Example 2: This invention also presents a schematic diagram of a computer device for identifying and detecting microcalcifications in mammograms based on deep learning. Please refer to [link / reference]. Figure 3 The computer device includes a memory 501, a processor 502, and a computer program 503 stored in the memory 501 and running on the processor 502. When the processor 502 executes the computer program 503, the computer device can execute any of the deep learning-based mammography image microcalcification point recognition and detection methods described above.

[0057] Furthermore, embodiments of this application also protect an apparatus that may include a memory and a processor, wherein the memory stores executable program code, and the processor is used to call and execute the executable program code to perform a deep learning-based method for identifying and detecting microcalcifications in mammography images provided in embodiments of this application.

[0058] This embodiment can divide the device into functional modules based on the above method example. For example, each module can correspond to a separate function, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware. It should be noted that the module division in this embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0059] When each module is divided according to its function, the device may also include a communication module, a signal analysis module, a complexity analysis module, and a positioning module. It should be noted that all relevant content of each step involved in the above method embodiments can be referenced from the functional descriptions of the corresponding functional modules, and will not be repeated here.

[0060] It should be understood that the apparatus provided in this embodiment is used to perform the above-described method for identifying and detecting microcalcifications in mammograms based on deep learning, and therefore can achieve the same effect as the above-described implementation method.

[0061] When using integrated units, the device may include a processing module and a storage module. When applied to a workpiece, the processing module can be used to control and manage the workpiece's operations. The storage module can be used to support the execution of program code by the workpiece.

[0062] The processing module may be a processor or a controller, which can implement or execute various exemplary logic blocks, modules, and circuits contained in conjunction with the disclosure of this application. The processor may also be a combination of functions that implement computing capabilities, such as a combination of one or more microprocessors, a combination of digital signal processing (DSP) and a microprocessor, etc., and the storage module may be a memory.

[0063] Example 3: This embodiment also provides a computer-readable storage medium storing computer program code. When the computer program code is run on a computer, the computer executes the above-described related method steps to implement the deep learning-based method for identifying and detecting microcalcifications in mammograms provided in the above embodiment.

[0064] Example 4: This embodiment also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned related steps to realize the deep learning-based method for identifying and detecting microcalcifications in mammograms provided in the above embodiment.

[0065] In this embodiment, the device, computer-readable storage medium, computer program product, or chip are all used to execute the corresponding methods provided above. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods provided above, and will not be repeated here.

[0066] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0067] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0068] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A method for identifying and detecting microcalcifications in mammograms based on deep learning, characterized in that, The method includes: Obtain mammograms of the test subjects; Several segmentation thresholds were determined; based on each segmentation threshold, the mammogram was segmented to identify the suspected microcalcification areas under each segmentation threshold. Based on the morphological, light and dark, and texture features of the suspected microcalcification point region, a feature vector is determined; based on the feature vector, a trained neural network is used to determine the confidence score of the suspected microcalcification point region. The confidence scores of each pixel in the suspected microcalcification region under different segmentation thresholds are corrected and fused using all segmentation thresholds to determine the final score of each pixel in the mammogram; based on the final score, the final microcalcification region is determined.

2. The method for identifying and detecting microcalcifications in mammograms based on deep learning according to claim 1, characterized in that, The determination of the feature vector includes: Calculate the standard deviation of the distances from the centroid of the suspected microcalcification region to all edge pixels, denoted as the shape dispersion; count the total number of pixels in the suspected microcalcification region, denoted as the region area; the morphological features include the shape dispersion and the region area. The brightness contrast is determined based on the grayscale difference between the suspected microcalcification area and its surrounding area; the brightness contrast is included in the light and dark features. The gradient values ​​of edge pixels in the suspected microcalcification region are obtained, and the average gradient values ​​of all edge pixels are calculated to obtain the edge saliency; the information entropy of the grayscale values ​​of pixels in the suspected microcalcification region is obtained and denoted as the texture saliency; the texture features include the edge saliency and the texture saliency. A feature vector is formed by the shape dispersion, area, brightness contrast, edge saliency, and texture saliency of the suspected microcalcification point region.

3. The method for identifying and detecting microcalcifications in mammography images based on deep learning according to claim 2, characterized in that, Determining the brightness contrast includes: For each edge pixel in the suspected calcification point region, draw a ray passing through the edge pixel with the centroid of the suspected microcalcification point region as the endpoint; extend the ray by a preset distance to obtain the analysis boundary point; Connect all edge pixels on the edge of the suspected microcalcification area to the corresponding analysis boundary points in sequence to form a closed boundary line; the area between the edge of the suspected microcalcification area and the boundary line is denoted as the background ring. The difference between the average gray value of pixels in the suspected microcalcification area and the average gray value of pixels in the background ring is calculated. The difference is used as the numerator, and the ratio obtained by dividing the average gray value of pixels in the background ring by a preset positive number is used as the denominator. This ratio is denoted as the brightness contrast ratio.

4. The method for identifying and detecting microcalcifications in mammograms based on deep learning according to claim 1, characterized in that, The determination of the final score for each pixel in the mammogram includes: Select a target threshold for each pixel in the mammogram, where the pixel is located within the suspected microcalcification area below the target threshold. The ratio of each target threshold for each pixel to the sum of all target thresholds is used as the threshold weight for that target threshold. Using the threshold weights of each target threshold for each pixel, the confidence scores of the pixel in the suspected microcalcification area under the corresponding target threshold are weighted and summed to obtain the final score of the pixel.

5. The method for identifying and detecting microcalcifications in mammograms based on deep learning according to claim 1, characterized in that, The determination of the final microcalcification point region includes: The connected region formed by pixels in the mammogram whose final score is greater than a preset first score threshold is denoted as the calcification core region. For each calcified core region, a morphological expansion operation is performed on the calcified core region to obtain an expanded region; the edge pixels of the calcified core region are used as the initial growth points, and the region is grown under the condition that the final score is greater than a preset second score threshold to obtain the final microcalcified point region.

6. The method for identifying and detecting microcalcifications in mammograms based on deep learning according to claim 1, characterized in that, The method for obtaining the segmentation threshold includes: The optimal threshold is obtained by using the maximum inter-class variance method for the gray values ​​of pixels in the mammogram. The optimal threshold is used as the minimum calcification threshold, and the maximum gray value of the pixels in the image is selected as the maximum calcification threshold. Within the grayscale range formed by the minimum calcification threshold and the maximum calcification threshold, scanning is performed with a preset step size to obtain different segmentation thresholds.

7. The method for identifying and detecting microcalcifications in mammograms based on deep learning according to claim 1, characterized in that, The suspected microcalcification region under each segmentation threshold is a connected region composed of pixels in the mammogram whose gray value is greater than or equal to the segmentation threshold.

8. The method for identifying and detecting microcalcifications in mammograms based on deep learning according to claim 5, characterized in that, The preset first score threshold is greater than the preset second score threshold.

9. The method for identifying and detecting microcalcifications in mammograms based on deep learning according to claim 1, characterized in that, The trained neural network is a fully connected neural network; The training process includes: obtaining feature vectors containing real microcalcifications and false positive samples, where false positive samples include malignant calcifications and vascular artifacts; and using the cross-entropy function as the loss function.

10. The method for identifying and detecting microcalcifications in mammograms based on deep learning according to claim 6, characterized in that, The preset step size is less than the difference between the maximum calcification threshold and the minimum calcification threshold.

Images