Bilateral breast volume difference estimation method for single-view point cloud chest shape reconstruction

Abstract

The invention discloses a bilateral breast volume difference estimation method for single-view point cloud chest shape reconstruction. The method comprises the following steps: constructing a chest simulation three-dimensional grid based on a chest shape model, constructing a simulation data set, constructing a neural network and training, predicting a real image, mapping the real image into a real world coordinate, and calculating the volume of a breast needing to be filled. Only one depth camera needs to be deployed to obtain the real image of the breast part of the patient, the interference to the operation environment is small, and the cost is low; on the basis of the neural network model trained by the constructed chest simulation data set, perception missing areas such as the back, the oxter, the nipple and the incision can be complemented, a three-dimensional grid describing complete shape information of the chest is obtained, and therefore the volume difference of the breasts on the left side and the right side can be accurately calculated; the volume needing to be filled after the single-side breast is cut off is accurately estimated, then an operation is guided to ensure that the breasts of a patient are symmetrical after the operation, and the operation quality is improved.

Description

Technical Field

[0001] The present invention belongs to the field of auxiliary medical care and intelligent medical care, and in particular to a method for estimating bilateral breast volume difference by single-view point cloud chest shape reconstruction. Background Art

[0002] With the improvement of people's health awareness and the continuous progress of clinical imaging technology, the early detection rate of breast cancer has approached 90%. At the same time, the prognosis of breast cancer patients has continued to improve, with a 5-year survival rate of over 90% and a 10-year survival rate of over 80%. At present, surgical methods that remove breast tumors but preserve the appearance of the breast (breast-conserving) are gradually becoming mainstream because they can significantly improve the quality of life of patients after surgery, such as traditional breast-conserving surgery, breast-reduction and breast-conserving surgery, and breast reconstruction surgery with autologous or prosthetic tissues.

[0003] As a special hanging organ, the shape of the breast will change due to gravity and changes in body position. During the operation, the patient lies on his back and his breast slides toward the chest wall. Therefore, it is necessary to evaluate the volume difference between the two sides of the breast to determine the volume that needs to be filled on the affected side, and then guide the operation to ensure that the patient's breasts are symmetrical after the operation.

[0004] The existing methods for evaluating the difference in bilateral breast volume are mainly divided into two categories: objective measurement and subjective visual inspection. Objective measurement includes mechanical measurement, magnetic resonance imaging, and 2D or 3D appearance imaging. However, mechanical measurement has the disadvantages of contact, inconvenient measurement, and difficulty in accurately describing non-rigid structural surfaces; magnetic resonance imaging relies on large equipment and complex image reconstruction; 2D and 3D imaging are mostly measured in a standing position, which is more complicated and time-consuming; therefore, the existing objective measurement methods are not suitable for use during surgery. For this reason, subjective visual inspection is basically used during surgery to evaluate the difference in bilateral breast volume, which relies on the doctor's visual experience and is easily affected by individual differences and visual errors. The postoperative symmetry can only reach 70%, making it difficult to provide accurate and reliable evaluation results. Summary of the invention

[0005] The present invention aims to solve the above-mentioned technical problems existing in the prior art and provides a method for estimating the volume difference of bilateral breasts by reconstructing the chest shape from a single-view point cloud.

[0006] The technical solution of the present invention is: a method for estimating the difference in bilateral breast volume by single-view point cloud chest shape reconstruction, which is performed in the following steps:

[0007] Step 1. Construct a 3D chest simulation mesh

[0008] Step 1.1 Build the basic chest model

[0009] The SMPL human body shape model is cut to keep the chest area, and the shape of the chest area is changed using the SMPL shape parameters to obtain basic chest models of different shapes.

[0010] Step 1.2 Basic chest model enhancement

[0011] Step 1.2.1: Divide the left and right parts of the front of the chest according to the vertex coordinate components x and z, that is, x>0, z>0 is the front left part, x<0, z<0 is the front right part, and take the basic chest models of the front left part and the front right part to form a chest model set in a 1:1 ratio;

[0012] Step 1.2.2: Assume that the left part of the front side is the cutting side and the right part of the front side is the healthy side, or the right part of the front side is the cutting side and the left part of the front side is the healthy side, and perform the following operations:

[0013] Step 1.2.2.1 Take the highest point on the cut side of the chest, take the highest point as the center of the circle, select an area with a radius of 0.01m, and subtract the random value d from the depth of this area. 1 ;

[0014] Step 1.2.2.2 Take the highest point on the cut side of the chest, take the highest point as the center of the circle, select an area with a radius of 0.015m, and subtract the random value d from the depth of this area. 2 ;

[0015] Step 1.2.2.3 Take the highest point on the cut side of the chest, take the highest point as the center of the circle, select an area with a radius of 0.025m, and subtract the random value d from the depth of this area. 3 ;

[0016] Step 1.2.2.4 Select the vertex with |x| < 0.025 on the healthy side of the chest and subtract d 2 , select the vertex with 0.0025≤|x|≤0.04, and subtract d 2 +0.001, imitates smooth transition;

[0017] Step 1.3 Select a reference point C between the back and front of the chest.

[0018] Point, calculate the difference between the vertex and point C, and get a series of direction vectors

[0019] Step 1.4 Calculate the direction vector The modulus length is used to determine the direction from point C to the apex of the chest on the cutting side;

[0020] Step 1.5 generates a random size value d, scales the original vertex position according to formula (1), and then obtains a simulated three-dimensional mesh;

[0021]

[0022] Where: V is the original vertex position, V′ is the scaled vertex position;

[0023] Step 2: Build a simulation dataset

[0024] Step 2.1 uses an orthogonal projection camera to orthogonally project the simulated triangular mesh into a depth map and store the human body shape parameters. During the projection process, each pixel only rasterizes one surface;

[0025] Step 2.2 performs data enhancement operation on the depth map;

[0026] Step 3: Build a neural network and train it

[0027] Step 3.1 Build a neural network

[0028] Construct a neural network consisting of a backbone module and a vertex coordinate prediction head module. The backbone module is composed of the convolution and residual parts of ResNet50. The vertex coordinate prediction head module is a MeshNet network, which is specifically provided with a deconvolution layer deconv, a one-dimensional convolution layer conv for predicting the x and y axis coordinates x 、conv y , convolutional layer conv used to predict the z-axis coordinate z1 、conv z2 ;

[0029] Step 3.2 Training

[0030] Step 3.2.1 Input the enhanced depth map into the backbone module to extract image features and obtain the feature map X′;

[0031] Step 3.2.2 Input the feature map X′ to the deconvolution layer deconv, adjust the dimension and then input it to the one-dimensional convolution layer conv x 、conv y , predict the x and y dimension values ​​coordx and coordy of the vertex coordinates, and perform the operation according to formulas (2) and (3);

[0032] coordx=soft_arg max(convx(avgx(deconv(X′)))) (2)

[0033] coordy=soft_arg max(convy(avgy(deconv(X′)))) (3)

[0034] Step 3.2.3 Input the feature map X′ to the convolutional layer conv z1, after dimension adjustment, it is input into the convolutional layer conv z2 , predict the z-dimension value coordz of the vertex coordinates, and perform the operation according to formulas (4) and (5);

[0035] featz=(convz 1 (avgz(deconv(X′)))) (4)

[0036] coordz = soft_arg max(conv z2 (fit(featz))) (5)

[0037] Step 3.2.4 uses formula (6) to integrate the corresponding values ​​of the three dimensions to obtain the predicted chest vertex coordinate coord;

[0038]

[0039] Step 3.2.5 generates real-world coordinates online based on the human body shape parameters stored when generating the depth map;

[0040] Step 3.2.6: Map the real world coordinates to the heat map size according to formulas (7) and (8);

[0041]

[0042] p″=p′×c (8)

[0043] Where: p is the value of the real point coordinate in one of the dimensions x, y, and z, p′ is the value after mapping to [0, 1], and p″ is the corresponding value of the heat map size coordinate. When p is the x or y dimension, a=200, b=400, c=96; when p is the z dimension, a=1, b=2, c=64;

[0044] Step 3.2.7 supervises the predicted chest vertex coordinate coord with the corresponding value p″ of the heat map size coordinate to obtain the predicted chest vertex heat map size coordinate p″′;

[0045] Step 3.2.8: Using the loss function L shown in formula (9), the optimizer is used to iteratively train the neural network to obtain a trained neural network model;

[0046]

[0047] Where: N is the number of input samples, y i is the value p″ corresponding to the coordinate of the heat map size, is the predicted chest vertex heat map size coordinate p″′;

[0048] Step 4. Predict the real image

[0049] The real chest cut image is input into the trained neural network model for processing to obtain the predicted real chest vertex heat map size coordinates p r ″ e ' al ;

[0050] Step 5. Mapping to real-world coordinates

[0051] According to formulas (10) and (11), the predicted real chest vertex heat map size coordinates p r ″ e ' al Mapping back to real-world coordinates;

[0052]

[0053] p o ″=p 0 ×ba (11)

[0054] Where: p o is the denormalized coordinate, p o ″ is the predicted real world coordinate of the chest vertex, when p o When x or y, a=200, b=400, c=96, when p o When it is z, a=1, b=2, c=64;

[0055] Step 6. Calculate breast volume to be filled

[0056] According to p o ″The positive and negative difference between the x-coordinates of all points and the x-values ​​of the center point coordinates, o "The breast is divided into two parts, left and right, and the volume is calculated using the convex hull method. The volume of the breast that needs to be filled is obtained based on the volume difference between the left and right sides.

[0057] The preferred step 2.2 performs data enhancement operation on the depth map, and is performed according to the following steps:

[0058] Step 2.2.1 Direction Enhancement

[0059] From [θ min ,θ max ] Random sampling in the interval θ n , construct the rotation matrix R according to formula (12) z (θ n );

[0060]

[0061] Afterwards, use the rotation matrix R z (θn ), rotate the left and right chest vertices of the depth map. The specific operation is shown in formula (13):

[0062] v i ′=R z (θ)·v i (13)

[0063] Where: v i is the point cloud before rotation, v i ′ is the point cloud after rotation;

[0064] Step 2.2.2 Random masking

[0065] Use the random mask to bitwise multiply the generated depth map, and the black area of ​​the random mask covers the generated depth map;

[0066] Step 2.2.3 Morphological corrosion

[0067] A randomly generated integer in the range of [1,10] is used as the erosion unit, and the erosion function is applied to perform morphological erosion on the masked depth map.

[0068] Compared with the prior art, the present invention has the following advantages: only one depth camera needs to be deployed to obtain a real image of the patient's breast area, which has little interference with the surgical environment and is low in cost; the neural network model trained based on the constructed chest simulation data set can complete the perceived missing areas such as the back, armpits, nipples, and incisions, and obtain a three-dimensional grid that describes the complete shape information of the chest, so as to accurately calculate the volume difference between the left and right breasts, and more accurately estimate the volume that needs to be filled after removing a single breast, thereby guiding the surgery to ensure the symmetry of the patient's breasts after surgery and improving the quality of the surgery. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 is a flow chart of an embodiment of the present invention.

[0070] Figure 2 A three-dimensional shape diagram of chest simulation data generated by an embodiment of the present invention.

[0071] Figure 3 A depth map of chest simulation data generated by an embodiment of the present invention.

[0072] Figure 4 Schematic diagram of the neural network structure of an embodiment of the present invention.

[0073] Figure 5 This is a projection image of a point predicted by an embodiment of the present invention on a depth map.

[0074] Figure 6 This is a three-dimensional shape diagram of the chest predicted by an embodiment of the present invention. DETAILED DESCRIPTION

[0075] The present invention provides a method for estimating the difference in bilateral breast volume based on single-view point cloud chest shape reconstruction. Figure 1 As shown, follow the steps below:

[0076] Step 1. Construct a 3D chest simulation mesh

[0077] Step 1.1 Build the basic chest model

[0078] The SMPL human body shape model is cropped to retain the chest area. The resulting chest mesh contains 600 vertices and 1084 triangles in total. The shape of the chest area is changed using SMPL shape parameters to obtain basic chest models of different shapes.

[0079] Step 1.2 Basic chest model enhancement

[0080] Step 1.2.1: Divide the left and right parts of the front of the chest according to the vertex coordinate components x and z, that is, x>0, z>0 is the front left part, x<0, z<0 is the front right part, and take the basic chest models of the front left part and the front right part to form a chest model set in a 1:1 ratio;

[0081] Step 1.2.2: Assume that the left part of the front side is the cutting side and the right part of the front side is the healthy side, or the right part of the front side is the cutting side and the left part of the front side is the healthy side, and perform the following operations:

[0082] Step 1.2.2.1 Take the highest point on the cut side of the chest, take the highest point as the center of the circle, select an area with a radius of 0.01m, and subtract the random value d from the depth of this area. 1 , d 1 ∈[0.008,0.01];

[0083] Step 1.2.2.2 Take the highest point on the cut side of the chest, take the highest point as the center of the circle, select an area with a radius of 0.015m, and subtract the random value d from the depth of this area. 2 , d 2 ∈[0.005,0.008];

[0084] Step 1.2.2.3 Take the highest point on the cut side of the chest, take the highest point as the center of the circle, select an area with a radius of 0.025m, and subtract the random value d from the depth of this area. 3 , d 3 ∈[0.006,0.009];

[0085] Step 1.2.2.4 Since the chest of the SMPL model is in the shape of a standing patient, and the patient is in a lying position during surgery, the chest naturally flows to both sides, resulting in a relatively flat center of the chest. At this time, the healthy side of the chest near the central axis is relatively convex, so the following treatment is performed:

[0086] Select the vertex with |x| < 0.025 on the healthy side of the chest, and subtract d 2 , select the vertex with 0.0025≤|x|≤0.04, and subtract d 2 +0.001, imitates smooth transition;

[0087] Step 1.3 Select a reference point C between the back and front of the chest.

[0088] Point, calculate the difference between the vertex and point C, and get a series of direction vectors

[0089] Step 1.4 Calculate the direction vector The modulus length is used to determine the direction from point C to the apex of the chest on the cutting side;

[0090] Step 1.5 generates a random size value d, and scales the original vertex position according to formula (1), thereby obtaining Figure 2 The simulated three-dimensional mesh shown;

[0091]

[0092] Where: V is the original vertex position, V′ is the scaled vertex position;

[0093] Step 2: Build a simulation dataset

[0094] Step 2.1 Use an orthogonal projection camera to orthogonally project the simulated triangular mesh into a depth map (the focal length defaults to 1) and store the human body shape parameters. During the projection process, each pixel only rasterizes one surface;

[0095] Step 2.2: Perform data enhancement on the depth map as follows:

[0096] Step 2.2.1 Direction Enhancement

[0097] From [θ min ,θ max ] Random sampling in the interval θ n , construct the rotation matrix R according to formula (2) z (θ n );

[0098]

[0099] Afterwards, use the rotation matrix Rz (θ n ), rotate the chest vertices of the left and right parts of the depth map. The specific operation is shown in formula (3):

[0100] v i ′=R z (θ)·v i (3)

[0101] Where: v i is the point cloud before rotation, v i ′ is the point cloud after rotation;

[0102] Step 2.2.2 Random masking

[0103] Use a random mask to multiply the generated depth map bit by bit, and the black area of ​​the random mask covers the generated depth map to simulate the holes in the real image caused by environmental interference;

[0104] Step 2.2.3 Morphological corrosion

[0105] A randomly generated integer in the range of [1,10] is used as the erosion unit, and the erosion function is applied to perform morphological erosion on the masked depth map, imitating the loss of edge information caused by the depth camera's insensitivity to edge perception, and removing small noise and small objects.

[0106] The depth map of the chest simulation data generated by the embodiment of the present invention is as follows: Figure 3 shown.

[0107] Step 3: Build a neural network and train it

[0108] Step 3.1 Build a neural network

[0109] Build as Figure 4 The neural network shown in the figure is composed of a backbone module and a vertex coordinate prediction head module, wherein the backbone module is composed of the convolution and residual parts of ResNet50, and the vertex coordinate prediction head module is a MeshNet network, which is specifically provided with a deconvolution layer deconv, a one-dimensional convolution layer conv for predicting the x and y axis coordinates x 、conv y , convolutional layer conv used to predict the z-axis coordinate z1 、conv z2 ;

[0110] Step 3.2 Training

[0111] Step 3.2.1 Input the enhanced depth map into the backbone module to extract image features and obtain the feature map X′;

[0112] Step 3.2.2 Input the feature map X′ to the deconvolution layer deconv, adjust the dimension and then input it to the one-dimensional convolution layer conv x 、conv y , predict the x and y dimension values ​​coordx and coordy of the vertex coordinates, and perform the operation according to formulas (4) and (5);

[0113] coordx=soft_arg max(convx(avgx(deconv(X′)))) (4)

[0114] coordy=soft_arg max(convy(avgy(deconv(X′)))) (5)

[0115] Step 3.2.3 Input the feature map X′ to the convolutional layer conv z1 , after dimension adjustment, it is input into the convolutional layer conv z2 , predict the z-dimension value coordz of the vertex coordinates, and perform the operation according to formulas (6) and (7);

[0116] featz=(convz 1 (avgz(deconv(X′)))) (6)

[0117] coordz = soft_arg max(conv z2 (fit(featz))) (7)

[0118] Step 3.2.4 uses formula (8) to integrate the corresponding values ​​of the three dimensions to obtain the predicted chest vertex coordinate coord;

[0119]

[0120] Step 3.2.5: Reduce the depth d according to the human body shape parameters stored when generating the depth map, such as the rotation matrix. 1 d 2 d 3 and parameters such as size reduction to generate real-world coordinates online;

[0121] Step 3.2.6: According to formulas (9) and (10), map the real world coordinates to the heat map size (96×96×64);

[0122]

[0123] p″=p′×c (10)

[0124] Where: p is the value of the real point coordinate in one of the dimensions x, y, and z, p′ is the value after mapping to [0, 1], and p″ is the corresponding value of the heat map size coordinate. When p is the x or y dimension, a=200, b=400, c=96; when p is the z dimension, a=1, b=2, c=64;

[0125] Step 3.2.7 supervises the predicted chest vertex coordinate coord with the corresponding value p″ of the heat map size coordinate to obtain the predicted chest vertex heat map size coordinate p″′;

[0126] Step 3.2.8 uses the loss function L shown in formula (11) and an optimizer to iteratively train the neural network to obtain a trained neural network model;

[0127]

[0128] Where: N is the number of input samples, y i is the value p″ corresponding to the coordinate of the heat map size, is the predicted chest vertex heat map size coordinate p″′;

[0129] The embodiment of the present invention uses the Adam optimizer for optimization. The constructed simulation data training set size is 2000, the test set size is 500, the image size is 369×369, the batch size is 8, the number of iterations is 13, and the initial learning rate is 10 -4 ;

[0130] Step 4. Predict the real image

[0131] The real chest cut image is input into the trained neural network model for processing to obtain the predicted real chest vertex heat map size coordinates p r ″ e ' al ;

[0132] Step 5. Mapping to real-world coordinates

[0133] According to formulas (12) and (13), the predicted real chest vertex heat map size coordinate p r ″ e ' al Mapping back to real-world coordinates;

[0134]

[0135] p o ″=p 0 ×ba (13)

[0136] Where: p o is the denormalized coordinate, po ″ is the real world coordinate of the predicted true chest vertex, when p o When x or y, a=200, b=400, c=96, when p o When it is z, a=1, b=2, c=64;

[0137] Visualize the projection of the predicted chest vertices on the depth map as follows Figure 5 As shown; the three-dimensional shape of the chest is Figure 6 As shown;

[0138] Step 6. Calculate breast volume to be filled

[0139] According to p o ″ The positive and negative difference between the x-coordinates of all points and the x-values ​​of the center point coordinates is p o ″ Divide it into left and right parts and calculate the volume using the convex hull method. Based on the volume difference between the left and right sides, the volume of the breast that needs to be filled is obtained.

Claims

1. A method for estimating bilateral breast volume difference based on single-view point cloud chest shape reconstruction, characterized in that Follow these steps: Step 1. Construct a 3D chest simulation mesh Step 1.1 Build the basic chest model The SMPL human body shape model is cut to keep the chest area, and the shape of the chest area is changed using the SMPL shape parameters to obtain basic chest models of different shapes. Step 1.2 Basic chest model enhancement Step 1.2.1: Divide the left and right parts of the front of the chest according to the vertex coordinate components x and z, that is, x>0, z>0 is the front left part, x<0, z<0 is the front right part, and take the basic chest models of the front left part and the front right part to form a chest model set in a 1:1 ratio; Step 1.2.2: Assume that the left part of the front side is the cutting side and the right part of the front side is the healthy side, or the right part of the front side is the cutting side and the left part of the front side is the healthy side, and perform the following operations: Step 1.2.2.1 Take the highest point on the chest on the cutting side, take the highest point as the center of the circle, select an area with a radius of 0.01m, and subtract the random value d1 from the depth of this area; Step 1.2.2.2 Take the highest point on the chest on the cutting side, take the highest point as the center, select an area with a radius of 0.015m, and subtract the random value d2 from the depth of this area; Step 1.2.2.3 Take the highest point on the chest on the cutting side, take the highest point as the center, select an area with a radius of 0.025m, and subtract the random value d3 from the depth of this area; Step 1.2.2.4 Select the vertices with |x|<0.025 on the healthy side chest, subtract d2, select the vertices with 0.0025≤|x|≤0.04, subtract d2+0.001, and simulate a smooth transition; Step 1.3 Select a reference point C between the back and front of the chest. For each vertex on the cut side, calculate the difference between the vertex and point C to obtain a series of direction vectors Step 1.4 Calculate the direction vector The modulus length is used to determine the direction from point C to the apex of the chest on the cutting side; Step 1.5 generates a random size value d, scales the original vertex position according to formula (1), and then obtains a simulated three-dimensional mesh; Where: V is the original vertex position, V′ is the scaled vertex position; Step 2: Build a simulation dataset Step 2.1 uses an orthogonal projection camera to orthogonally project the simulated triangular mesh into a depth map and store the human body shape parameters. During the projection process, each pixel only rasterizes one surface; Step 2.2 performs data enhancement operation on the depth map; Step 3: Build a neural network and train it Step 3.1 Build a neural network Construct a neural network consisting of a backbone module and a vertex coordinate prediction head module. The backbone module is composed of the convolution and residual parts of ResNet50. The vertex coordinate prediction head module is a MeshNet network, which is specifically provided with a deconvolution layer deconv, a one-dimensional convolution layer conv for predicting the x and y axis coordinates x 、conv y , convolutional layer conv used to predict the z-axis coordinate z1 、conv z2 ; Step 3.2 Training Step 3.2.1 Input the enhanced depth map into the backbone module to extract image features and obtain the feature map X′; Step 3.2.2 Input the feature map X′ to the deconvolution layer deconv, adjust the dimension and then input it to the one-dimensional convolution layer conv x 、conv y , predict the x and y dimension values ​​coordx and coordy of the vertex coordinates, and perform the operation according to formulas (2) and (3); coordx=soft_arg max(convx(avgx(deconv(X′)))) (2) coordy=soft_arg max(convy(avgy(deconv(X′)))) (3) Step 3.2.3 Input the feature map X′ to the convolutional layer conv z1 , after dimension adjustment, it is input into the convolutional layer conv z2 , predict the z-dimension value coordz of the vertex coordinates, and perform the operation according to formulas (4) and (5); featz=(convz1(avgz(deconv(X′)))) (4) coordz=soft_arg max(conv z2 (fit(featz))) (5) Step 3.2.4 uses formula (6) to integrate the corresponding values ​​of the three dimensions to obtain the predicted chest vertex coordinate coord; Step 3.2.5 generates real-world coordinates online based on the human body shape parameters stored when generating the depth map; Step 3.2.6: Map the real world coordinates to the heat map size according to formulas (7) and (8); p″=p′×c (8) Where: p is the value of the real point coordinate in one of the dimensions x, y, and z, p′ is the value after mapping to [0, 1], and p″ is the corresponding value of the heat map size coordinate. When p is the x or y dimension, a=200, b=400, c=96; when p is the z dimension, a=1, b=2, c=64; Step 3.2.7 supervises the predicted chest vertex coordinate coord with the corresponding value p″ of the heat map size coordinate to obtain the predicted chest vertex heat map size coordinate p″′; Step 3.2.8: Using the loss function L shown in formula (9), the optimizer is used to iteratively train the neural network to obtain a trained neural network model; Where: N is the number of input samples, y i is the value p″ corresponding to the coordinate of the heat map size, is the predicted chest vertex heat map size coordinate p″′; Step 4. Predict the real image The real chest cut image is input into the trained neural network model for processing to obtain the predicted real chest vertex heat map size coordinates p r ″ e ' al ; Step 5. Mapping to real-world coordinates According to formulas (10) and (11), the predicted real chest vertex heat map size coordinates p r ″ e ' al Mapping back to real-world coordinates; p o ″=p0×b-a (11) Where: p o is the denormalized coordinate, p o ″ is the predicted real world coordinate of the chest vertex, when p o When x or y, a=200, b=400, c=96, when p o When it is z, a=1, b=2, c=64; Step 6. Calculate breast volume to be filled According to p o ″The positive and negative difference between the x-coordinates of all points and the x-values ​​of the center point coordinates, o "The breast is divided into two parts, left and right, and the volume is calculated using the convex hull method. The volume of the breast that needs to be filled is obtained based on the volume difference between the left and right sides.

2. The method for estimating bilateral breast volume difference by single-view point cloud breast shape reconstruction according to claim 1, characterized in that The step 2.2 performs data enhancement operation on the depth map according to the following steps: Step 2.2.1 Direction Enhancement From [θ min ,θ max ] Random sampling in the interval θ n , construct the rotation matrix R according to formula (12) z (θ n ); Afterwards, use the rotation matrix R z (θ n ), rotate the left and right chest vertices of the depth map. The specific operation is shown in formula (13): v i ′=R z (θ)·v i (13) Where: v i is the point cloud before rotation, v i ′ is the point cloud after rotation; Step 2.2.2 Random masking Use the random mask to bitwise multiply the generated depth map, and the black area of ​​the random mask covers the generated depth map; Step 2.2.3 Morphological corrosion A randomly generated integer in the range of [1,10] is used as the erosion unit, and the erosion function is applied to perform morphological erosion on the masked depth map.

Images