A method and system for assisted identification of a breast mass region

Abstract

The present application relates to the technical field of image processing and recognition, and particularly relates to a breast lump region auxiliary recognition method and system. According to the difference between the gradient distribution in the neighborhood range of each edge pixel point in any two overlapping prediction boxes and the direction difference between the center point and the edge pixel point distribution of each prediction box, the centripetal degree of each prediction box relative to the center point is obtained. According to the difference between the centripetal degree of each prediction box relative to the center point and the center point of another prediction box in any two overlapping prediction boxes, the lump difference degree of the corresponding two prediction boxes is obtained. The intersection over union weighted function adjusted by the lump difference degree is used to suppress the non-maximum value of the prediction box, and the target box in the mammogram image is screened. Based on the target box, the breast lump region in the mammogram image is recognized. The breast lump feature difference of the overlapping prediction box is used to avoid excessive suppression of the confidence degree, and the accuracy of the breast lump recognition is improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and specifically to an auxiliary identification method and system for breast lump regions. Background Technology

[0002] Breast cancer is one of the most common malignant tumors in women. Early screening and diagnosis are crucial for improving the cure rate of breast cancer. Mammography is the recognized standard screening method for early diagnosis of breast cancer. With the rapid development of artificial intelligence technology, computer-aided diagnostic systems based on deep learning have been widely used for the automated detection of breast lumps, especially target detection algorithms, which have made significant progress in the automatic detection of breast lump areas in X-ray images.

[0003] Soft Non-Maximum Suppression (Soft-NMS) is a bounding box determination method in object detection algorithms. The Soft-NMS algorithm relies on the intersection-union ratio between detection boxes to adjust the confidence of overlapping boxes. However, when multiple detection boxes cover the same tumor area, the Soft-NMS algorithm often ignores the detailed information of the area within the detection box, resulting in the over-suppression of the confidence of the detection boxes in the overlapping area, thus missing the detection of breast mass areas. Summary of the Invention

[0004] To address the technical problem of missed detection of breast lumps due to excessive suppression of the confidence level of the detection frame, the present invention aims to provide an auxiliary identification method and system for breast lumps, the specific technical solution of which is as follows: In a first aspect, one embodiment of the present invention provides an auxiliary identification method for a breast lump region, the method comprising: Obtain the predicted bounding box in a mammogram image; Based on the gradient distribution of edge pixels within the neighborhood of each edge pixel in any two overlapping prediction boxes, and the directional difference between the center point of each prediction box and the distribution of each edge pixel, obtain the centripetal degree of each prediction box relative to its center point in any two overlapping prediction boxes. Based on the difference in centripetal intensity between any two overlapping prediction boxes relative to their center point and the center point of the other prediction box, the mass difference degree between any two overlapping prediction boxes is obtained. Using the intersection-union weighted function in the soft non-maximum suppression algorithm after the mass difference adjustment, non-maximum suppression is applied to the predicted bounding box to filter target boxes in the mammogram image; based on the target boxes, the breast mass region in the mammogram image is identified.

[0005] Further, obtaining the centripetal degree of each prediction box relative to its center point in any two overlapping prediction boxes includes: Based on the gradient distribution of edge pixels within the neighborhood of each edge pixel in each prediction box, obtain the analysis gradient vector of each edge pixel in each prediction box; The direction of each edge pixel in each prediction box pointing to its center point is taken as the direction of the centripetal vector of the corresponding edge pixel, and the distance of each edge pixel in each prediction box pointing to its center point is taken as the magnitude of the centripetal vector of the corresponding edge pixel. The mean of the negative correlation mapping results of the angle between the analytical gradient vector and the centripetal vector of all edge pixels in the region where each prediction box does not contain the other prediction box in any two overlapping prediction boxes is calculated to obtain the centripetality of each prediction box relative to its center point in the corresponding two prediction boxes.

[0006] Further, obtaining the analysis gradient vector of each edge pixel within each prediction box includes: Edge detection is performed on the mammogram to obtain the edge line of each prediction box; for each prediction box, a reference point is determined for each edge pixel within the prediction box, wherein the reference point is within the neighborhood of each edge pixel and the reference point is on the same edge line as each edge pixel. Calculate the proportion of the distance between each edge pixel and each reference point within the prediction box to the sum of the distances between each edge pixel and all its reference points, and use the difference between the constant 1 and the proportion as the reference weight of each reference point for the corresponding edge pixel. Obtain the gradient vector of each edge pixel within the prediction box; add the gradient vector of the reference point of each edge pixel within the prediction box to the reference weight to obtain the analysis gradient vector of the corresponding edge pixel.

[0007] Further, obtaining the tumor dissimilarity of any two overlapping prediction boxes includes: Obtain the centripetal value of each prediction box relative to the center point of the other prediction box in any two overlapping prediction boxes; The sum of the absolute values ​​of the centripetal differences between each prediction box relative to its center point and the center point of the other prediction box in any two overlapping prediction boxes is normalized to obtain the tumor difference degree of the corresponding two overlapping prediction boxes.

[0008] Furthermore, the method of screening target boxes in mammogram images includes: The confidence score of each predicted bounding box in a mammogram image is obtained using an object detection algorithm. An initial detection box is determined, and the remaining predicted boxes are designated as candidate boxes. The set of detection boxes is designated as the target set. The intersection-union ratio (IUR) between the detection boxes and each candidate box is obtained. It is determined whether the IUR is greater than a preset IUR threshold. If it is, the IUR weighting function is adjusted using the tumor difference between the detection boxes and the candidate boxes. The confidence of the candidate boxes is then weighted using the adjusted IUR weighting function to obtain the updated confidence of the candidate boxes. Otherwise, the confidence of the candidate boxes remains unchanged. Based on the updated confidence, candidate boxes that do not meet the selection criteria are filtered out. A new detection box is selected from the remaining candidate boxes and added to the target set. The remaining candidate boxes, excluding the new detection boxes, are filtered out again until there are no remaining candidate boxes. The predicted boxes in the updated target set are designated as target boxes.

[0009] Furthermore, the formula for calculating the updated confidence score of the candidate box is as follows: , In the formula, Let be the updated confidence level of the a-th candidate box; Let M be the confidence score of the a-th candidate box; M is the detection box. is the intersection-union ratio (IoU) between the detected bounding box and the a-th candidate bounding box; Yiou is the preset IoU threshold. The cross-union ratio weighted function is the tumor difference adjustment function between the detected bounding box and the a-th candidate box; The difference between the detection box and the a-th candidate box is the degree of difference in the mass. is a hyperparameter; exp is an exponential function with the natural constant as the base.

[0010] Furthermore, the candidate condition is that the updated confidence level of the candidate box is greater than a preset confidence threshold.

[0011] Further, obtaining the centripetal degree of each prediction box relative to the center point of the other prediction box in any two overlapping prediction boxes includes: The direction of each edge pixel in each prediction box pointing to the center point of the other prediction box is taken as the direction of the non-self-centripetal vector of the corresponding edge pixel, and the distance between each edge pixel in each prediction box and the center point of the other prediction box is taken as the magnitude of the non-self-centripetal vector of the corresponding edge pixel. The mean of the negative correlation mapping results of the angle between the analysis gradient vector and the non-self centripetal vector of all edge pixels in the region where each prediction box does not contain the other prediction box in any two overlapping prediction boxes is calculated to obtain the centripetal degree of each prediction box relative to the center point of the other prediction box.

[0012] Furthermore, the method for obtaining the neighborhood range includes: Edge detection is performed on mammogram images to obtain edge pixels; For each predicted bounding box in a mammogram image, the confidence level of the predicted bounding box, the difference between the constant 1 and the confidence level of the predicted bounding box are used as the weights of the preset maximum and preset minimum values ​​of the neighborhood radius, respectively. The preset maximum and preset minimum values ​​of the neighborhood radius are then weighted and summed to obtain the preferred neighborhood radius of the edge pixels within the predicted bounding box. The region centered on each edge pixel within the prediction box and with the preferred neighborhood radius as its radius is defined as the neighborhood range of each edge pixel within the prediction box.

[0013] Secondly, another embodiment of the present invention provides an auxiliary identification system for breast lump regions, the system comprising: The data acquisition module is used to acquire the prediction bounding box in the mammogram image; The centripetal analysis module is used to obtain the centripetal degree of each prediction box relative to its center point in any two overlapping prediction boxes based on the gradient distribution of edge pixels within the neighborhood of each edge pixel in each prediction box in any two overlapping prediction boxes and the directional difference between the center point of each prediction box and the distribution of each edge pixel. The mass difference analysis module is used to obtain the mass difference degree of any two overlapping prediction boxes based on the difference in centripetal degree of each prediction box relative to its center point and the center point of the other prediction box. The breast mass auxiliary identification module is used to perform non-maximum suppression on the predicted bounding box using the intersection-union weighted function in the soft non-maximum suppression algorithm after the mass difference adjustment, and to filter the target bounding box in the mammogram image; and to identify the breast mass region in the mammogram image based on the target bounding box.

[0014] The present invention has the following beneficial effects: In this embodiment of the invention, based on the characteristic that the gradient direction of the edge of a breast lump points to the center of the lump, the centripetal force of the edge pixels within the predicted bounding box is analyzed by the directional difference between the gradient distribution of the edge pixels and the distribution of the edge pixels and the center point, thus measuring the probability that the edge pixels within the predicted bounding box belong to the breast lump within the box. The gradient direction of the edge pixels within two predicted bounding boxes covering different breast lump regions should point to different lump centers. By analyzing the difference in centripetal force between each predicted bounding box relative to its center point and the center point of the other predicted bounding box in two overlapping predicted bounding boxes, the lump difference between the two predicted bounding boxes is analyzed, thus measuring the possibility that the two predicted bounding boxes represent different breast lump regions. Based on the lump difference between different predicted bounding boxes, the confidence suppression process of the traditional Soft-NMS algorithm, i.e., the cross-union weighting function, is adjusted to avoid two overlapping predicted bounding boxes belonging to different breast lumps being over-suppressed, effectively improving the accuracy of the confidence adjustment of the predicted bounding boxes, reducing the risk of misjudgment of overlapping bounding boxes and loss of breast lumps, and improving the accuracy of breast lump screening and the efficiency of clinical diagnosis. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the steps of an auxiliary identification method for breast lump regions provided in one embodiment of the present invention; Figure 2 This is a flowchart illustrating the steps of a method for obtaining the difference in mass size according to an embodiment of the present invention. Figure 3 This is a system structure diagram of an auxiliary identification system for breast lump regions provided in one embodiment of the present invention. Detailed Implementation

[0016] The following description, in conjunction with the accompanying drawings, details the specific scheme of the auxiliary identification method and system for breast lump regions provided by the present invention.

[0017] Example 1: This invention proposes an auxiliary method for identifying breast lump regions. Please refer to [link / reference]. Figure 1 The diagram illustrates a flowchart of a method for assisting in the identification of a breast lump region according to an embodiment of the present invention. The method includes: Step S1: Obtain the prediction box in the mammogram image.

[0018] Specifically, the female patient being tested stands or sits in front of a mammography machine after removing her top and bra. The doctor points the machine at her breast and takes an X-ray image. This image is then fed into a target detection algorithm, which outputs predicted bounding boxes and the confidence score for each box. Since the breast of a breast cancer patient is mainly composed of adipose tissue, glandular tissue, and breast lumps, the object categories identified by the target detection algorithm in the mammogram image include: breast lumps, adipose tissue, and glandular tissue. The purpose of this approach is to identify breast lumps; the confidence score of a predicted bounding box indicates the probability that the box contains a breast lump.

[0019] In this embodiment of the invention, the target detection algorithm is the YOLO algorithm. In other embodiments, the target detection algorithm may also be Faster Region-based Convolutional Neural Networks (Faster R-CNN), Mask Region-based Convolutional Neural Networks (Mask R-CNN), RetinaNet, etc.

[0020] Step S2: Based on the gradient distribution of edge pixels within the neighborhood of each edge pixel in any two overlapping prediction boxes, and the directional difference between the center point of each prediction box and the distribution of each edge pixel, obtain the centripetal degree of each prediction box relative to its center point in any two overlapping prediction boxes.

[0021] The breast is mainly composed of adipose tissue, glandular tissue, and breast masses. Adipose tissue absorbs less X-rays, making it appear as a darker area on X-ray images. Breast masses are tumorous or fibrous tissue proliferations, typically denser than normal glandular tissue, and therefore absorb more X-rays, resulting in a higher grayscale value than glandular tissue on X-ray images. In summary, breast masses exhibit a high grayscale characteristic, and the grayscale value gradually decreases as pixels move away from the mass, eventually returning to the grayscale of normal glandular tissue. Since the gradient change is strongest at the edge of the breast mass, the gradient direction at the edge usually points towards the center of the mass. The prediction box is used to locate the breast mass in a mammogram; typically, the center point of the prediction box can be considered the possible center of the breast mass. Therefore, the direction of the edge pixels pointing towards the center of the mass is close to the gradient direction of the edge pixels.

[0022] The gradient distribution of each edge pixel within the neighborhood of the predicted bounding box reflects the gradient direction of each edge pixel. The difference between the gradient distribution of each edge pixel and the center point of the predicted bounding box and the distribution direction of each edge pixel is compared, and the centripetal degree of each edge pixel relative to its center point is considered.

[0023] Preferably, in some possible implementations of the embodiments of the present invention, the method for obtaining centripetality includes: obtaining the analysis gradient vector of each edge pixel in each prediction box based on the gradient distribution of edge pixels within the neighborhood of each edge pixel in each prediction box; taking the direction of each edge pixel pointing to its center point in each prediction box as the direction of the centripetal vector of the corresponding edge pixel, and taking the distance of each edge pixel pointing to its center point in each prediction box as the magnitude of the centripetal vector of the corresponding edge pixel; calculating the mean of the negative correlation mapping results of the angle between the analysis gradient vector and the centripetal vector of all edge pixels in the region where each prediction box does not contain the other prediction box in any two overlapping prediction boxes, to obtain the centripetality of each prediction box relative to its center point in the two corresponding prediction boxes.

[0024] Considering that the edge of a breast lump is not necessarily a smooth circle or ellipse, and that small blood vessels near the lump edge can cause the edge gradient to be not perpendicular to the actual lump edge, leading to errors in subsequent centripetal analysis, in order to suppress the interference caused by factors such as blood vessels near the lump edge, the gradient information of the edge pixels in the neighborhood of each edge pixel is used to smooth the edge gradient, resulting in the analysis gradient vector of the edge pixels within the prediction box.

[0025] In some possible implementations of this invention, the method for obtaining the gradient vector analysis includes: performing edge detection on a mammogram to obtain the edge line of each prediction box; for each prediction box, determining a reference point for each edge pixel within the prediction box, wherein the reference point is within the neighborhood of each edge pixel and lies on the same edge line as each edge pixel; calculating the proportion of the distance between each edge pixel and each reference point within the prediction box in the sum of the distances between each edge pixel and all its reference points, and using the difference between the constant 1 and the proportion as the reference weight for each reference point of the corresponding edge pixel; obtaining the gradient vector of each edge pixel within the prediction box; and weightedly adding the gradient vector of the reference point of each edge pixel within the prediction box to obtain the analysis gradient vector of the corresponding edge pixel. Specifically, the scalar multiplication result of the gradient vector of the reference point of each edge pixel within the prediction box and the reference weight is used as the weighted result, and vector addition is performed on the scalar multiplication results corresponding to all reference points of the edge pixel to obtain the analysis gradient vector of that edge pixel. It should be noted that the closer the edge pixel is to its reference point, the higher the reference value of the gradient information of the reference point, and the greater the reference weight. The distance between two pixels refers to the Euclidean distance; the edge line of the prediction box can be obtained by curve fitting the edge pixels within the prediction box; the curve fitting method is the least squares method, which is a well-known technique to those skilled in the art and will not be described in detail here.

[0026] In this embodiment of the invention, the method for obtaining the neighborhood range is as follows: edge detection is performed on the mammogram image to obtain edge pixels; for each predicted box in the mammogram image, the confidence level of the predicted box, the difference between the constant 1 and the confidence level of the predicted box are respectively used as the weights of the preset maximum value and preset minimum value of the neighborhood radius, and the preset maximum value and preset minimum value of the neighborhood radius are weighted and summed to obtain the preferred neighborhood radius of the edge pixels within the predicted box; the region formed with each edge pixel within the predicted box as the center and the preferred neighborhood radius as the radius is taken as the neighborhood range of each edge pixel within the predicted box. It should be noted that the higher the confidence level of the predicted box, the higher the probability of the presence of a breast mass, and a larger neighborhood range can be assigned to the edge pixels within the predicted box; conversely, a smaller neighborhood range is assigned to the lower confidence level to prevent the edge gradient of the mass from being overly smoothed. In this embodiment of the invention, the formula for calculating the preferred neighborhood radius of each edge pixel within the predicted box is: In the formula, The preferred neighborhood radius of the edge pixel within the i-th prediction box; Let be the confidence score of the i-th predicted bounding box; Set a maximum value for the neighborhood radius; A minimum value is preset for the neighborhood radius; The rounding up symbol.

[0027] In this embodiment of the invention, the neighborhood radius is preset to a maximum value. With the preset minimum value of the neighborhood radius The spatial resolution of mammogram images and the prior morphological scale of the breast mass were used as the basis for the design. It should be larger than the size of isolated noise points to suppress local gradient perturbations, and smaller than the minimum effective radius of a typical mass to avoid over-smoothing of edge features; preferably set to 9 pixels. Set to 3 to ensure that there are enough pixels in the neighborhood for gradient direction statistics.

[0028] In one implementation of this invention, the Canny operator is used to perform edge detection on mammogram images.

[0029] Assuming the y-th and z-th predicted boxes overlap, taking the y-th and z-th predicted boxes as an example, the centripetal degree of the y-th predicted box relative to its center point within the y-th and z-th predicted boxes is... Expressed as a formula: In the formula, The total number of edge pixels in the region where the z-th prediction box is not included in the y-th prediction box; The analysis gradient vector is the nth edge pixel within the region that does not contain the zth prediction box in the yth prediction box; Let be the centripetal vector of the nth edge pixel within the region that does not contain the zth prediction box in the yth prediction box; cos is the cosine function, which calculates the negative correlation mapping of the angle between the gradient vector and the centripetal vector. It is important to note that when prediction boxes completely overlap or completely contain each other, the total number of edge pixels within the region that does not contain another prediction box is considered. If the value is zero, the centripetal value of the corresponding predicted box is directly set to zero. The y-th and z-th predicted boxes may represent the same breast mass, but there are differences between them. Whether two predicted boxes represent the same breast mass depends mainly on the non-overlapping portion between them. Therefore, in the analysis... The process only requires analyzing the regions in the y-th prediction box that do not contain the z-th prediction box. .like The smaller the angle between the analysis gradient vector of the inner edge pixel and the centripetal vector, the greater the probability that the inner edge pixel of the y-th predicted bounding box points to the center of the breast mass. The larger it is, the more it indicates The higher the probability that an inner edge pixel belongs to a breast mass within the y-th prediction box.

[0030] Step S3: Based on the difference in centripetal intensity of each prediction box relative to its center point and the center point of the other prediction box in any two overlapping prediction boxes, obtain the tumor difference degree of any two overlapping prediction boxes.

[0031] If two predicted bounding boxes cover the same breast lump region, the pixel gradient changes at the lump edges are consistent, and the lump edges converge in the direction of the lump center. This means that the gradient directions of the edge pixels within both predicted boxes will also tend to point towards the lump center. If the two predicted bounding boxes cover different breast lump regions, the gradient direction of the edge pixels within each predicted box should point towards the lump center represented by the predicted box, meaning the gradient directions of the edge pixels within the two predicted boxes point to different lump centers. Therefore, by analyzing the difference in centripetal intensity between the two predicted boxes relative to their center points and the center point of the other predicted box, we can assess the likelihood that the two predicted boxes represent different breast lump regions.

[0032] Please see Figure 2 The diagram illustrates a flowchart of a method for obtaining the difference in mass size according to an embodiment of the present invention. The method includes: Step S310: Obtain the centripetal value of each prediction box relative to the center point of the other prediction box in any two overlapping prediction boxes.

[0033] The specific method is as follows: take the direction of each edge pixel in each prediction box pointing to the center point of the other prediction box in any two overlapping prediction boxes as the direction of the non-self-centripetal vector of the corresponding edge pixel, and take the distance between each edge pixel in each prediction box and the center point of the other prediction box as the magnitude of the non-self-centripetal vector of the corresponding edge pixel; calculate the mean of the negative correlation mapping results of the angle between the analysis gradient vector and the non-self-centripetal vector of all edge pixels in the region of each prediction box that does not contain the other prediction box in any two overlapping prediction boxes, and obtain the centripetal degree of each prediction box relative to the center point of the other prediction box in the corresponding two prediction boxes.

[0034] Centimeter of the y-th prediction box relative to the center point of the z-th prediction box within the y-th and z-th prediction boxes Expressed as a formula: In the formula, The total number of edge pixels in the region where the z-th prediction box is not included in the y-th prediction box; The analysis gradient vector is the nth edge pixel within the region that does not contain the zth prediction box in the yth prediction box; is the non-self-centripetal vector of the nth edge pixel in the region where the zth prediction box is not included in the yth prediction box; cos is the cosine function, which realizes the negative correlation mapping of the angle by calculating and analyzing the cosine value of the angle between the gradient vector and the non-self-centripetal vector.

[0035] Step S320: Normalize the sum of the absolute values ​​of the differences in centripetal values ​​between each prediction box relative to its center point and the center point of the other prediction box in any two overlapping prediction boxes to obtain the tumor difference degree of the corresponding two overlapping prediction boxes.

[0036] The difference in mass distribution between the y-th and z-th predicted bounding boxes Expressed as a formula: In the formula, Let y be the centripetal value of the y-th prediction box relative to its center point; Let y be the centripetal value of the y-th prediction box relative to the center point of the z-th prediction box; The centripetal value of the z-th prediction box relative to the center point of the y-th prediction box; Let z be the centripetal value of the z-th prediction box relative to its center point; The absolute value function is used; Norm is the normalization function. Since centripetal force is represented by the mean of the cosine of the included angle, its range is... Therefore, the maximum absolute value of any two centripetal differences is 2, and the theoretical maximum value of the sum of the absolute values ​​of the two differences is 4. Based on the theoretical maximum value, the maximum value normalization method is used to... Normalization was performed. It should be noted that... and The analysis object is the edge pixels within the region where the z-th prediction box does not contain the y-th prediction box; if and The larger the value, the greater the probability that the gradient direction of the edge pixels within the y-th and z-th predicted boxes points to different tumor centers, and the greater the likelihood that the y-th and z-th predicted boxes represent different breast tumor regions. This indicates a higher tumor dissimilarity. The larger.

[0037] It is important to note that and , and The method for obtaining the values ​​is similar; simply replace the region in the y-th prediction box that does not contain the z-th prediction box with the region in the z-th prediction box that does not contain the y-th prediction box. All other aspects remain the same. Obtaining the tumor difference between any two overlapping prediction boxes is done in the same way as obtaining the tumor difference between the y-th and z-th prediction boxes.

[0038] Step S4: Using the intersection-union weighted function in the soft non-maximum suppression algorithm after mass difference adjustment, non-maximum suppression is applied to the predicted bounding box to filter the target bounding box in the mammogram image; the breast mass region in the mammogram image is identified based on the target bounding box.

[0039] The Soft-NMS algorithm is suitable for handling overlapping targets and densely occluded scenes. It suppresses the confidence of overlapping predicted boxes, ultimately obtaining predicted boxes with higher confidence to improve detection accuracy and recall. However, when multiple detection boxes cover the same tumor region, the Soft-NMS algorithm often ignores the detailed information within the detection boxes, leading to excessive suppression of the confidence of detection boxes in overlapping areas. This solution adjusts the confidence suppression process of the traditional Soft-NMS algorithm based on the tumor dissimilarity of different predicted boxes, thereby selecting target boxes.

[0040] Preferably, in some possible implementations of the embodiments of the present invention, the method for obtaining the target bounding box includes: obtaining the confidence score of each predicted bounding box in a mammogram image using a target detection algorithm; determining an initial detection bounding box, recording the remaining predicted bounding boxes as candidate bounding boxes, recording the set composed of the detection bounding boxes as the target set, obtaining the intersection-union ratio (IU) between the detection bounding box and each candidate bounding box, determining whether the IU is greater than a preset IU threshold, if so, adjusting the IU weighting function using the mass difference between the detection bounding box and the candidate bounding box, weighting the confidence scores of the candidate bounding boxes using the adjusted IU weighting function to obtain the updated confidence scores of the candidate bounding boxes, otherwise, keeping the confidence scores of the candidate bounding boxes unchanged; filtering out candidate bounding boxes that do not meet the selection criteria based on the updated confidence scores, selecting new detection bounding boxes from the remaining candidate bounding boxes, adding the new detection bounding boxes to the target set, continuing to filter out the remaining candidate bounding boxes except for the new detection bounding boxes until there are no remaining candidate bounding boxes, and recording the predicted bounding boxes in the updated target set as target bounding boxes. The Soft-NMS algorithm is a well-known technology to those skilled in the art and will not be described in detail here.

[0041] It should be noted that the initial detection box is the predicted box corresponding to the highest confidence level in the mammogram image, and the new detection box is the predicted box corresponding to the highest confidence level among the candidate boxes. The selection criterion is that the updated confidence level of the candidate box is greater than a preset confidence threshold.

[0042] The formula for calculating the updated confidence score of the candidate boxes is: In the formula, For the first The updated confidence of each candidate box; For the first The confidence scores of each candidate bounding box; M represents the detection bounding box; For the detection box and the first The intersection-union ratio of candidate boxes; Yiou is the preset intersection-union threshold; The cross-union ratio weighted function is the tumor difference adjustment function between the detected bounding box and the k-th candidate bounding box; For the detection box and the first The difference in mass distribution among candidate boxes; For hyperparameters, an empirical value of 0.4 is used in this embodiment; exp is an exponential function with the natural constant as the base. It should be noted that in the traditional Soft-NMS algorithm, the detection box and the... The expression for the intersection-union ratio weighted function of the candidate boxes is: .because ,but Two overlapping detection frames belonging to different breast lumps and the first Each candidate box will not be excessively suppressed; and if The larger the detection box is, the closer it is to the first... The greater the likelihood that each candidate box represents a different breast mass region, the more attention needs to be paid to the detailed information of the regions within the two predicted boxes, and consequently, the less likely it is to suppress the first candidate box. The confidence level of each candidate box. The larger, the more The larger.

[0043] In another embodiment of the invention, after the detection frame and the first The expression for the cross-union ratio weighted function after adjusting for the malignancy differences of each candidate box can also be: .

[0044] In this embodiment of the invention, a preset intersection-union threshold is used to determine whether two predicted boxes have significant spatial overlap, and a preset confidence threshold is used to filter out candidate boxes with too low confidence after adjustment for mass difference. Specifically, the acquisition method is as follows: collect several historical mammogram images and their corresponding labeled mass regions; use a target detection algorithm to obtain the predicted boxes and their confidence scores in each sample image, and calculate the intersection-union ratio (IUR) between each predicted box and the labeled ground truth box; statistically analyze the IUR distribution characteristics of all predicted boxes, and set the preset IUR threshold to a specific percentile value in the IUR distribution, such as the 85th percentile; statistically analyze the confidence distribution characteristics of each predicted box after non-maximum suppression, and set the preset confidence threshold to a specific percentile value in the confidence distribution, such as the 30th percentile.

[0045] As an example, if a mammogram image has prediction boxes a, b, c, d, and e, the confidence levels are as follows: , , , , ,in If the value is the largest, then the predicted box a is denoted as the initial detection box, the predicted boxes b, c, d, and e are denoted as candidate boxes, and the set consisting of the detection boxes a is denoted as the target set. Calculate the intersection-union ratios (IURs) of detection box a with candidate boxes b, c, d, and e, respectively. Assuming that the IURs of detection box a with candidate boxes b, c, and d are all greater than a preset IU threshold, and the IUR of detection box a with candidate box e is less than the preset IU threshold, then the confidence level of candidate box e is... Without weighting, the above method is used to predict bounding boxes b, c, and d. , , After weighting, the updated confidence levels of predicted boxes b, c, and d are as follows: , , Filter out , , , Candidate boxes with confidence scores less than or equal to a preset confidence threshold are assumed to be... If the threshold is less than or equal to the preset threshold, then delete e, and the remaining candidate boxes are b, c, and d.

[0046] like , , middle If the predicted bounding box is the largest, then the predicted bounding box b is used as the new detection bounding box, and the remaining candidate bounding boxes are c and d. The new detection bounding box b is then added to the target set. In the process, the updated target set is obtained. Calculate the intersection-union ratios (IoU) of detection box b with candidate boxes c and d, respectively. Assuming that the IoU of detection box b with candidate boxes c and d are both greater than a preset IoU threshold, use the above method to predict the boxes c and d. , After weighting, the updated confidence levels of predicted boxes c and d are as follows: , Filter out , Candidate boxes with confidence scores less than or equal to a preset confidence threshold are assumed to be... , If both predicted boxes are less than or equal to the preset confidence threshold, then predicted boxes c and d are deleted, leaving no candidate boxes. At this point, the target set is... The predicted bounding boxes a and b are denoted as the target boxes.

[0047] It should be noted that this scheme modifies the intersection-union (IU) weighting function in the Soft-NMS algorithm by adjusting the IU weighting function based on the tumor difference between the detection box and the candidate box. The adjusted IU weighting function is then used to weight the confidence of the candidate box, while all other aspects remain unchanged. If the detection box and the candidate box do not overlap, the confidence of the candidate box remains unchanged.

[0048] Target bounding boxes with a confidence level greater than a preset mass threshold in mammogram images are selected as suspected mass bounding boxes, indicating a higher probability of a breast mass within the suspected mass bounding box. Corresponding rectangular regions are drawn on the mammogram images based on the suspected mass bounding boxes. These rectangular regions are the initially identified breast mass regions. The final confirmation is made by combining the doctor's professional knowledge and experience, assisting the doctor in completing the regional identification and annotation of breast masses, and providing visualization results of mass boundary information, including the size, location, and malignancy risk probability assessment of the mass.

[0049] In one implementation of this invention, a preset tumor threshold is used to determine whether the target box constitutes a valid breast tumor region, and can be set to an empirical value of 0.5. Alternatively, several historical mammogram images of labeled breast tumor regions are collected, and a target detection algorithm is used to obtain the confidence values ​​of all target boxes in each sample. The distribution characteristics of the confidence values ​​are statistically analyzed, and the preset tumor threshold is set to a specific percentile value in the confidence distribution, such as the 50th percentile.

[0050] This solution improves upon the traditional Soft-NMS algorithm by leveraging the gradient characteristics of the mass edges in X-ray images. This effectively enhances the accuracy of confidence adjustment for predicted bounding boxes, reduces the risk of misjudgment of overlapping bounding boxes and loss of breast masses, and improves the accuracy of breast mass screening and the efficiency of clinical diagnosis.

[0051] This invention is now complete.

[0052] Example 2: This invention proposes an auxiliary identification system for breast lump regions. Please refer to [link / reference]. Figure 3 The diagram illustrates a system structure of an auxiliary identification system for breast lump regions according to an embodiment of the present invention. The system includes: Data acquisition module 510 is used to acquire prediction boxes in mammogram images; The centripetal analysis module 520 is used to obtain the centripetal degree of each prediction box relative to its center point in any two overlapping prediction boxes based on the gradient distribution of edge pixels within the neighborhood of each edge pixel in each prediction box in any two overlapping prediction boxes and the directional difference between the center point of each prediction box and the distribution of each edge pixel. The mass difference analysis module 530 is used to obtain the mass difference degree of any two overlapping prediction boxes based on the difference in centripetal degree of each prediction box relative to its center point and the center point of another prediction box. The breast mass auxiliary identification module 540 is used to perform non-maximum suppression on the predicted bounding box using the intersection-union weighted function in the soft non-maximum suppression algorithm after mass difference adjustment, and to filter the target bounding box in the mammogram image; and to identify the breast mass region in the mammogram image based on the target bounding box.

[0053] It should be noted that the devices provided in the above embodiments are only illustrative examples of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the auxiliary identification system for breast lump regions and the auxiliary identification method for breast lump regions provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

Claims

1. A method for assisting in the identification of breast lump regions, characterized in that, The method includes: Obtain the predicted bounding box in a mammogram image; Based on the gradient distribution of edge pixels within the neighborhood of each edge pixel in any two overlapping prediction boxes, and the directional difference between the center point of each prediction box and the distribution of each edge pixel, obtain the centripetal degree of each prediction box relative to its center point in any two overlapping prediction boxes. Based on the difference in centripetal intensity between any two overlapping prediction boxes relative to their center point and the center point of the other prediction box, the mass difference degree between any two overlapping prediction boxes is obtained. Using the intersection-union weighted function in the soft non-maximum suppression algorithm after the mass difference adjustment, non-maximum suppression is applied to the predicted bounding box to filter target boxes in the mammogram image; based on the target boxes, the breast mass region in the mammogram image is identified.

2. The method for auxiliary identification of breast lump regions according to claim 1, characterized in that, The step of obtaining the centripetal value of each prediction box relative to its center point in any two overlapping prediction boxes includes: Based on the gradient distribution of edge pixels within the neighborhood of each edge pixel in each prediction box, obtain the analysis gradient vector of each edge pixel in each prediction box; The direction of each edge pixel in each prediction box pointing to its center point is taken as the direction of the centripetal vector of the corresponding edge pixel, and the distance of each edge pixel in each prediction box pointing to its center point is taken as the magnitude of the centripetal vector of the corresponding edge pixel. The mean of the negative correlation mapping results of the angle between the analytical gradient vector and the centripetal vector of all edge pixels in the region where each prediction box does not contain the other prediction box in any two overlapping prediction boxes is calculated to obtain the centripetality of each prediction box relative to its center point in the corresponding two prediction boxes.

3. The method for auxiliary identification of breast lump regions according to claim 2, characterized in that, The step of obtaining the analysis gradient vector of each edge pixel within each prediction box includes: Edge detection is performed on the mammogram to obtain the edge line of each prediction box; for each prediction box, a reference point is determined for each edge pixel within the prediction box, wherein the reference point is within the neighborhood of each edge pixel and the reference point is on the same edge line as each edge pixel. Calculate the proportion of the distance between each edge pixel and each reference point within the prediction box to the sum of the distances between each edge pixel and all its reference points, and use the difference between the constant 1 and the proportion as the reference weight of each reference point for the corresponding edge pixel. Obtain the gradient vector of each edge pixel within the prediction box; add the gradient vector of the reference point of each edge pixel within the prediction box to the reference weight to obtain the analysis gradient vector of the corresponding edge pixel.

4. The method for auxiliary identification of breast lump regions according to claim 2, characterized in that, The step of obtaining the tumor difference between any two overlapping prediction boxes includes: Obtain the centripetal value of each prediction box relative to the center point of the other prediction box in any two overlapping prediction boxes; The sum of the absolute values ​​of the centripetal differences between each prediction box relative to its center point and the center point of the other prediction box in any two overlapping prediction boxes is normalized to obtain the tumor difference degree of the corresponding two overlapping prediction boxes.

5. The method for auxiliary identification of breast lump regions according to claim 1, characterized in that, The target bounding box for screening mammogram images includes: The confidence score of each predicted bounding box in a mammogram image is obtained using an object detection algorithm. An initial detection box is determined, and the remaining predicted boxes are designated as candidate boxes. The set of detection boxes is designated as the target set. The intersection-union ratio (IUR) between the detection boxes and each candidate box is obtained. It is determined whether the IUR is greater than a preset IUR threshold. If it is, the IUR weighting function is adjusted using the tumor difference between the detection boxes and the candidate boxes. The confidence of the candidate boxes is then weighted using the adjusted IUR weighting function to obtain the updated confidence of the candidate boxes. Otherwise, the confidence of the candidate boxes remains unchanged. Based on the updated confidence, candidate boxes that do not meet the selection criteria are filtered out. A new detection box is selected from the remaining candidate boxes and added to the target set. The remaining candidate boxes, excluding the new detection boxes, are filtered out again until there are no remaining candidate boxes. The predicted boxes in the updated target set are designated as target boxes.

6. The method for auxiliary identification of breast lump regions according to claim 5, characterized in that, The formula for calculating the updated confidence level of the candidate box is: , In the formula, Let be the updated confidence level of the a-th candidate box; Let M be the confidence score of the a-th candidate box; M is the detection box. is the intersection-union ratio (IoU) between the detected bounding box and the a-th candidate bounding box; Yiou is the preset IoU threshold. The cross-union ratio weighted function is the tumor difference adjustment function between the detected bounding box and the a-th candidate box; The difference between the detection box and the a-th candidate box is the degree of difference in the mass. is a hyperparameter; exp is an exponential function with the natural constant as the base.

7. The method for auxiliary identification of breast lump regions according to claim 5, characterized in that, The candidate condition is that the updated confidence level of the candidate box is greater than a preset confidence threshold.

8. The method for auxiliary identification of breast lump region according to claim 4, characterized in that, The process of obtaining the centripetal value of each prediction box relative to the center point of the other prediction box in any two overlapping prediction boxes includes: The direction of each edge pixel in each prediction box pointing to the center point of the other prediction box is taken as the direction of the non-self-centripetal vector of the corresponding edge pixel, and the distance between each edge pixel in each prediction box and the center point of the other prediction box is taken as the magnitude of the non-self-centripetal vector of the corresponding edge pixel. The mean of the negative correlation mapping results of the angle between the analysis gradient vector and the non-self centripetal vector of all edge pixels in the region where each prediction box does not contain the other prediction box in any two overlapping prediction boxes is calculated to obtain the centripetal degree of each prediction box relative to the center point of the other prediction box.

9. The method for auxiliary identification of breast lump regions according to claim 1, characterized in that, The method for obtaining the neighborhood range includes: Edge detection is performed on mammogram images to obtain edge pixels; For each predicted bounding box in a mammogram image, the confidence level of the predicted bounding box, the difference between the constant 1 and the confidence level of the predicted bounding box are used as the weights of the preset maximum and preset minimum values ​​of the neighborhood radius, respectively. The preset maximum and preset minimum values ​​of the neighborhood radius are then weighted and summed to obtain the preferred neighborhood radius of the edge pixels within the predicted bounding box. The region centered on each edge pixel within the prediction box and with the preferred neighborhood radius as its radius is defined as the neighborhood range of each edge pixel within the prediction box.

10. An auxiliary identification system for breast lump areas, characterized in that, The system includes: The data acquisition module is used to acquire the prediction bounding box in the mammogram image; The centripetal analysis module is used to obtain the centripetal degree of each prediction box relative to its center point in any two overlapping prediction boxes based on the gradient distribution of edge pixels within the neighborhood of each edge pixel in each prediction box in any two overlapping prediction boxes and the directional difference between the center point of each prediction box and the distribution of each edge pixel. The mass difference analysis module is used to obtain the mass difference degree of any two overlapping prediction boxes based on the difference in centripetal degree of each prediction box relative to its center point and the center point of the other prediction box. The breast mass auxiliary identification module is used to perform non-maximum suppression on the predicted bounding box using the intersection-union weighted function in the soft non-maximum suppression algorithm after the mass difference adjustment, and to filter the target bounding box in the mammogram image; and to identify the breast mass region in the mammogram image based on the target bounding box.

Images