LEARNING-BASED ACTIVE SURFACE MODEL FOR MEDICAL IMAGE SEGMENTATION

Abstract

A learning-based active surface model for medical image segmentation uses a method including: (a) data generation: obtaining medical images and associated ground truths, and splitting the sample images into a training set and a testing set; (b) raw segmentation: constructing a surface initialization network, parameters of the network trained by images and labels in the training set; (c) surface initialization: segmenting the images by the surface initialization network, and generating the point cloud data as the initial surface from the segmentation; (d) fine segmentation: constructing the surface evolution network, the parameters of the network trained by the initial surface obtained in step (c); (e) surface evolution: deforming the initial surface points along the offsets to obtain the predicted surface, the offsets presenting the prediction of the surface evolution network; (f) surface reconstruction: reconstructing the 3D volumes from the set of predicted surface points set to obtain the final segmentation results.

Claims

1. A learning-based active surface model for medical image segmentation, using a method comprising: step 1: data generation: obtaining the medical images and the associated ground truths, and splitting the sample images into a training set and a testing set; step 2: raw segmentation: constructing the surface initialization network, parameters of the network trained by images and labels in the training set; step 3: surface initialization: segmenting the images by surface initialization network, and generating the point cloud data as the initial surface from the segmentation; step 4: constructing the surface evolution network, the parameters of the network trained by the initial surface obtained in step 3; step 5: surface evolution: deforming the initial surface points along the offsets to obtain the predicted surface, the offsets presenting the prediction of the surface evolution network; step 6: surface reconstruction: reconstructing the 3D volumes from the set of predicted surface points set to obtain the final segmentation results.

2. The learning-based active surface model for medical image segmentation according to claim 1, wherein headstream part of the surface initialization network contains a convolutional layer, a batch normalization, an activation function and a max pool layer, then there are four residual modules; in the first two residual modules, the last convolutional layers have down-sampling operations, and the convolutional layers in the third and the fourth residual modules are dilated convolutions; it follows the atrous spatial pyramid pooling module consisting of a 1 × 1 convolutional layer, three dilated convolution layers, and a global pooling layer in parallel; another two 1 × 1 convolutions are used to integrate the features and predict a small resolution segmentation; finally, an up-sampling operation is used to restore the small resolution segmentation to the original image size.

3. The learning-based active surface model for medical image segmentation according to claim 2, wherein each pixel-wise label, obtained in step 1, is a binary image with the same size as the corresponding sample image, with value 1 indicating the desired target organ, and value 0 otherwise, the interpolation method used in up-sampling operation is bilinear interpolation.

4. The learning-based active surface model for medical image segmentation according to claim 1, wherein the training method of the surface initialization network in step 2; first, cut each sample image in the training set into two-dimensional slices along the third dimension; assuming the size of a three-dimensional CT image is H×W×D, the image can be regarded as two-dimensional images of size H×W stacked together, and each two-dimensional image is called a slice of a three-dimensional CT image; input the slices to the surface initialization network to obtain segmented predictions, where the binary cross-entropy is used as the loss function for training; the parameters of the initialization network are updated by the stochastic gradient descent method; the pixel-wise coarse segmentation predictions are obtained as follows; cut each sample image in the training set and testing set into two-dimensional slices along the third dimension; then sequentially input the slices of each three-dimensional CT image to the surface initialization network to obtain the two-dimensional segmentation predictions; finally stack the two-dimensional segmentations sequentially to obtain the corresponding three-dimensional coarse segmentation prediction.

5. The learning-based active surface model for medical image segmentation according to claim 1, wherein a method to pick up the surface points from the coarse segmentation is selecting the pixels of value 1, and at least one of their adjacent pixels is of value 0; let the set of selected surface points be S= {x.sub.1, x.sub.2, x.sub.3, ..., x.sub.n}, where n is the number of surface points; the narrow band of the surface S̃ = {x̃.sub.i ∈ P|d(x̃.sub.i,S) ≤ 1} = {x̃.sub.1,x̃.sub.2,x̃.sub.3,.sup....,x̃.sub.i,.sup....,x̃.sub.N} is used as the initial surface, where P represents all pixels of the $\text{and}d\left({\tilde{x}}_i,S\right)=\underset{x\in S}{\min }dist\left({\tilde{x}}_i,x\right)$ presents the distance from the pixel x̃.sub.i to the surface S image, and defined as the minimum distance between the pixel x̃.sub.i and the points in the surface points set S; the distance dist(x̃.sub.i, x) = ||x̃.sub.i - x||.sup.2 is the Euclidean distance between the two points x̃.sub.i and x, and N denotes the number of points in the set S̃; the set S̃ is used as the initial surface for evolution.

6. The learning-based active surface model for medical image segmentation according to claim 1, wherein the headstream part of the surface evolution network consists of a KP convolutional layer, a one-dimensional normalization and a leaky activation function; then there are five parts in the encoder of the network; the first and the last part contains two KP residual modules each, while the other three parts contain 3 KP residual modules each; the last KP residual modules in the first four parts have down-sampling operation; there are four up-sampling modules in the decoder of the network, each up-sampling module contains an up-sampling layer and a shared multilayer perceptron, followed by a normalization layer and an activation function; finally, two multilayer perceptrons are used to output the final prediction.

7. The learning-based active surface model for medical image segmentation according to claim 6, wherein some points are dropped in the down-sampling operations, and the up-sampling layers are used to restore those points by the nearest interpolation algorithm; after each up-sampling operation, the features are concatenated with the corresponding low-level features in the encoder, then the shared multilayer perceptron is used to aggregate the features; the inputs of the KP convolution are the point cloud data consisting of all points in a sphere centered on X̃, which is defined as.sup.N(x̃) = {z.sub.i ∈ P|dist(z.sub.i,x̃) ≤ r}, x̃ ∈ S̃, with r representing the radius of the sphere, X̃ presenting the point in the set of initial surface points, P representing all the pixels of the image, z.sub.i being the points in the sphere, and S̃ being the set of initial surface points; the KP convolution is an operation used to integrate the features; let denote the input features and ƒ.sub.i be the feature of the point z.sub.i; we define the kernel point convolution as $\left(F\ast g\right)\left(\tilde{x}\right)={\displaystyle \underset{z_i\in N\left(\tilde{x}\right)}{.Math.}g\left(z_i-\tilde{x}\right)}{f}_i,$ and the kernel function g as $\vartheta \left(z_i-\right)={\displaystyle \underset{h<K}{.Math.}\max \left(0,1-\frac{‖(z_i-\tilde{x})-y_k‖}{\sigma }\right)}{W}_k,$ where {y.sub.k|||y.sub.k||.sup.2 ≤ r, k < K} denotes the K kernel points, σ denotes the influence distance of the kernel points and {W.sub.k|k < K} is the associated weight matrixes to learn from the point cloud data.

8. The learning-based active surface model for medical image segmentation according to claim 7, wherein the following L2 norm between the prediction and target offsets is used as the loss function for the surface evolution network $L_{top-k}\left(\theta \right)=\frac{1}{z_i\in M}{\displaystyle \underset{z_i\in N\left(\tilde{x}\right)}{.Math.}\left({‖\delta \left(z_i\right)9\gamma \left(z_n\right)‖}^2\right)},$ where the loss function L.sub.top_k (θ) only supervises on the top k loss of convolutional points to focus on hard points, Z.sub.i represents all the convolutional points, δ(z.sub.i) are the ground-truth offsets of Z.sub.i, and γ(z.sub.i) are the corresponding predicted offsets; the stochastic gradient descent method is used to update the parameters θ of the network.

9. The learning-based active surface model for medical image segmentation according to claim 8, wherein estimate the position of the predicted surface points by evolving the initial surface points along the corresponding predicted offsets: $u_i={\tilde{x}}_i+\gamma \left({\tilde{x}}_i\right),\forall {\tilde{x}}_i\in \tilde{S},$ where x̂.sub.i presents the initial surface points, γ(x̃.sub.i) presents the corresponding predicted offsets of x̂.sub.i, and S̃ is the set of initial surface points.

10. The learning-based active surface model for medical image segmentation according to claim 8, wherein first use the Delaunay triangulation to build up the triangle mesh from the prediction surface points set U = {u.sub.1, u.sub.2, ..., u.sub.N}, and the alpha shape mesh is obtained by removing the tetrahedrons with a circumradius larger than 5; then the subdivide algorithm is used to voxelize the alpha shape mesh to obtain a closed surface, and finally adapt the flood fill algorithm to fill the holes in the closed surface and generate the final 3D segmentation.

11. The learning-based active surface model for medical image segmentation according to claim 3, wherein assuming that the size of a three-dimensional CT image is H×W×D, it is regarded as D two-dimensional images of size H×W stacked together; input the slices to the surface initialization network to obtain the predicted segmentations; calculate the difference between the predicted segmentations and the labels of the corresponding slices, where the binary cross-entropy is adapted as the loss function; the parameters of the surface initialization network are updated by the stochastic gradient descent method; the coarse segmentation predictions are obtained as follows: firstly, all the three-dimensional CT images in the training set and the testing set are cut into two-dimensional slices along the third dimension; then sequentially input the slices of each three-dimensional CT image to the surface initialization network to obtain the two-dimensional segmentation predictions; finally stack the two-dimensional segmentations sequentially to obtain the corresponding three-dimensional coarse segmentation prediction; that is, stacking D two-dimensional segmentations of size H×W into a three-dimensional image of size H× W×D, thus, the coarse segmentation is obtained.

12. The active surface model for medical image segmentation based on deep learning according to claim 5, wherein the headstream part of the surface evolution network consists of a KP convolutional layer, a one-dimensional normalization and a leaky activation function; then five parts are included in the encoder of the network; the first and the last part contain two KP residual modules each, while the others contain 3 KP residual modules each; the last KP residual module in the first four parts has the down-sampling operation; there are four up-sampling modules in the decoder of the network, each of which contains an up-sampling layer and a shared multilayer perceptron, followed by a normalization layer and an activation function; finally, two multilayer perceptrons are used to output the final prediction.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0030] To describe the embodiments and the technical solutions of the present invention clearly, the following will briefly introduce the figures. Whilst the following description describes exemplary embodiments, it will be understood by those skilled in the art that many variations of the embodiment can be made within the scope and spirit of the present invention.

[0031] FIG. 1 is the flow chart of the present invention.

[0032] FIG. 2 is the illustration of the surface initialization network structure of the invention.

[0033] FIG. 3 is the illustration of the surface evolution network structure of the invention.

[0034] FIG. 4 is the illustration of the KP convolution of the invention.

[0035] FIG. 5 is the comparison of two typical results between the present invention and other methods on the MSD spleen dataset.

[0036] FIG. 6 is the comparison of two typical results between the present invention and other methods on the public spleen dataset.

DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] The technical solutions for the embodiments are described below clearly with some figures. Obviously, the following description describes exemplary embodiments, it will be understood by those skilled in the art that many variations of the embodiment can be made within the scope and spirit of the present invention.

[0038] As shown in FIG. 1, the learning-based active surface model proposes a fully automated organ segmentation model for 3D CT scans. The organ segmentation model is inspired by a variational method, the active contour/surface, including two stages: surface initialization and surface evolution. In the first stage, obtain the coarse segmentation through the surface initialization network; then estimate the initial surface from the coarse segmentation. In the second stage, input the initial surface, a point cloud data, to the surface evolution network and obtain the offsets, then deform the initial surface according to the offsets. The final segmentation results are reconstructed from the evolved surface points. Specific steps are as follows:

[0039] Step 1: data generation: obtaining the medical images and the associated ground truths, and splitting the sample images into a training set and a testing set;

[0040] We obtain the three-dimensional CT medical images, and then split the images into two parts: training set and testing set, where the label data are given for data in the training set. The associated label is an image of the same size, whose pixels are marked by 0 or 1. Usually, the labeling work is done by professionals such as doctors.

[0041] Step 2: raw segmentation: constructing the surface initialization network, parameters of the network trained by images and labels in the training set;

[0042] As shown in FIG. 2, the structure of the initialization network is given as follows. The headstream part of the surface initialization network contains a convolutional layer, a batch normalization, an activation function, and a max pooling layer. Specifically, the kernel size and step size of the convoluitoinal layer is 7×7 and 4, respectively. The convolutional operations are linear transformations, while the activation function is a non-linear projection. The max pooling layer is a kind of down-sampling operation, which reduces the resolution of the images to remove redundant information and reduce calculations. Then there are 4 residual modules. The last convolution layers of the first two residual modules have a step size of 2. The convolutional layer with a step size of 2 is also a kind of down-sampling operation. The convolutional layers of the third residual module and the fourth residual module are dilated convolutions, the rates of the dilated convolutions are 2 and 4, respectively. The basic difference between the dilated convolution and the standard convolution is the kernel. There are some holes in the kernel of the dilated convolution, and the dilated rate controls the size of the holes. By the kernel with holes, the receptive field are enlarged without increasing parameters. The following is the atrous spatial pyramid pooling module, which consists of a 1×1 convolutional, three dilated convolutions, and a global pooling in parallel. The dilated rate of the three expanded convolutional layers are 6, 12, and 18 respectively. The atrous spatial pyramid pooling module is used to extract more potential the features. The output features of these parallel layers are concatenated together, and another two 1×1 convolutional layers are used to integrate these features and produce a small resolution segmentation.

[0043] Finally, the up-sampling operation with bilinear interpolation is used to restore the binary segmentation to the original size.

[0044] Train the parameters of the initialization network by the images and labels in the training set. First, we cut the images into two-dimensional slices along the third dimension. Assuming the size of a three-dimensional CT image is H×W×D, the image can be regarded as two-dimensional images of size H×W stacked together, and each two-dimensional image is called a slice of a three-dimensional CT image. We input the slices to the surface initialization network to obtain the corresponding segmentations, where the binary cross-entropy between the labels and predictions is used as the loss function for training. The parameters of the initialization network are updated by the stochastic gradient descent method.

[0045] Step 3: surface initialization: segmenting the images by surface initialization network, and generating the point cloud data as the initial surface from the segmentation;

[0046] Once the initial network is trained, a coarse segmentation model of CT images is obtained. The images of both the training and testing set are input to the initial network, and the network produces the coarse segmentation results. The coarse segmentation of the training set is used to train the surface evolution network, and the coarse segmentation of the testing set is used to test the surface evolution network in the next phase. The specific operations are as follows:

[0047] Firstly, we cut each sample image in the training set and testing set into two-dimensional slices along the third dimension. Then we sequentially input the slices of each three-dimensional CT image to the surface initialization network to obtain the two-dimensional segmentation predictions. Finally we stack the two-dimensional segmentations sequentially to obtain the corresponding three-dimensional coarse segmentation prediction. Specifically, we stack D two-dimensional segmentations of size H×W into a three-dimensional image of size H×W×D. However, the segmentation predicted by the surface initialization network is not desirable. Based on the coarse segmentation, some methods use another network to predict again. In contrast, the presented model uses a network to deform the inaccurate surface to the accurate object surface.

[0048] Estimating the surface means picking up the surface points from the coarse segmentation. Indeed, the value of the surface initial network prediction ranges in [0,1]. A threshold being 0.5 is adapted to transform the predicted probability map to a binary segmentation, with each pixel of value 0 or 1, where 1 represents the target object, and 0 otherwise. Each pixel has 6 adjacent positions of front and back, left and right, and up and down. A pixel is regarded as a surface point, if it is of value 1, and at least one of the six adjacent pixels is of value 0. To be noticed, the surface is not satisfactory, due to the unsatisfactory coarse segmentation.

[0049] Let the set of selected surface points be S = {x.sub.1,x.sub.2,x.sub.3,...,x.sub.n}, where n is the number of surface points. The narrow band of the surface S̃ = {x̃.sub.1, ∈ P|d(x̃.sub.i,S) ≤ 1} ={x̃.sub.1,x̃.sub.2,x̃.sub.3,...,x̃.sub.i,...,x̃.sub.N} is used as the initial surface, where P represents all pixels of the image, and d(x̃.sub.i, S) =

$d\left({\tilde{x}}_i,S\right)=\underset{x\in S}{\min }dist\left({\tilde{x}}_i,x\right)$

*.sub.6*' (iiR.sup.t(.sup.!.i, x) presents the distance from the pixel x̃.sub.i to the surface S defined as the minimum distance between the pixel x̃.sub.i and the points in the surface points set S. The distance dist(x̃.sub.i, x) = ∥x̃.sub.i - x∥.sup.2 is the Euclidean distance between the two points x̃.sub.i and x, and N denotes the number of points in the set ̃S̃. The set S̃ is used as the initial surface for evolution. Experimentally, using the point set of narrow band, rather than the set of selected surface points, as initial surface can increase the robustness of initialization and stability of the evolution.

[0050] On the other hand, the corresponding ground truth offsets are generated for training the surface evolution network according to the labels in the training set. Similarly, the pixel-wise label is binary with a value of 1 indicates the target organ, and 0 otherwise. The surface points are selected by the value of pixels and the adjacent pixel in the aforementioned way. Assuming the ground truth surface points set T contains m points T = {y.sub.1, y.sub.2, y.sub.3, ..., y.sub.m}, the ground truth offsets of initialization surface points x̃.sub.i. are uniquely determined by the δ(x̃.sub.i) = x̃.sub.i - y*, i = 1, 2, ...N, where x̃.sub.i presents the points in the initialization surface points set, and

$y*=arg\underset{y\in T}{\min }{‖{\tilde{x}}_i-y‖}^2$

presents the point in the ground truth surface points set whose distance to the point x̃.sub.i is minimal.

[0051] The ground truth offsets δ(x̃.sub.i) is the target to be predicted by the surface evolution network. And the initial surface in the training set and the corresponding ground truth offsets are training pairs for surface evolution network.

[0052] Step 4: fine segmentation: constructing the surface evolution network, the parameters of the network trained by the initial surface obtained in step 3;

[0053] Since the surface points are point cloud data in space, the present invention uses the point cloud network to learn the offsets from the data, rather than rearranging the surface points in space into a two-dimensional plane. Since the position information of the surface points destroys during the rearrangement, the two-dimensional network learning from the plane data, cannot leverage the position information of the surface points.

[0054] The surface evolution network leverages the KP convolution (kernel point convolution), which combines local features according to the 3D geometric structure of the point cloud data. The inputs of the KP convolution are point cloud data consisting of all neighboring points in a sphere centered on X̃, denoted as N(x̃) = {z.sub.i ∈ P|dist(z.sub.i, x̃) ≤ r}, x̃ ∈ S̃, where r represents the radius, x̃ presents the initial surface point, P represents all the pixels of the image, z.sub.i are the points in the sphere, and S̃ is the set of initial surface. Let F denote the input features and fi be the features of the point .sup.Zi. As shown in FIG. 4, kernel point convolution which associates the features with the kernel function .sup.g is defined as follows

$\left(F\ast g\right)\left(\tilde{x}\right)={\displaystyle \underset{z_i\in N\left(\tilde{x}\right)}{.Math.}g\left(z_i-\tilde{x}\right)}{f}_i,$

and the kernel function g is defined as:

$g\left(z_i-\tilde{x}\right)={\displaystyle \underset{k<K}{.Math.}\max \left(0,1-\frac{‖\left(z_i-\tilde{x}\right)-y_k‖}{\sigma }\right)}{W}_k,$

where {yk|∥yk||.sup.2 ≤ r,k < K} denotes the K kernel points, σ denotes the influence distance of the kernel points and {W.sub.k|k < K} is the associated weight matrixes to learn from the samples.

[0055] As shown in FIG. 3, The structure of the surface evolution network is as follows: the headstream part of the surface evolution network consists of a KP convolutional layer, a one-dimensional normalization and a leaky activation function. Then there are five parts in the encoder of the network. The first and the last part contains two KP residual modules, while the other three parts contain 3 KP residual modules. The last KP residual modules in the first four parts have a down-sampling operation, which randomly drops out some points to reduce the density of the points. The radius of the kernel point convolution are increased as the density decrease to enlarge the receptive fields. There is no down-sampling operation in the last part, since the non-local features extraction is efficient. Then there are four up-sampling modules, each up-sampling module contains an up-sampling layer and a shared multilayer perceptron, followed by a normalization layer and an activation function. The up-sampling operations in the up-sampling layers restore the dropped points by the nearest interpolation algorithm, such that the number of the prediction points is consistent with the number of the input. After each up-sampling operation, the features are concatenated with the corresponding low-level features in the encoder, then the shared multilayer perceptron is used to aggregate the features. Finally, there are two multilayer perceptrons and a one-dimensional normalization, which are used to output the prediction γ(x). The prediction is the offsets of the corresponding surface points.

[0056] Calculate the two norms of the prediction offsets of the network and the ground-truth offsets:

$L_{top\_k}\left(\theta \right)=\frac{1}{M}{\displaystyle \underset{z_i\in N\left(\tilde{x}\right)}{.Math.}\left({‖\delta \left(z_i\right)-\gamma \left(z_i\right)‖}^2\right)},$

where the loss function

$L_{top\_k}\left(\theta \right)$

"-IOP-,4-(O) only supervises on the top M loss of points to focus on hard samples , Z.sub.i represents the convolutional points, δ(zi) are the corresponding groundtruth offsets of Zi, γ(zi) is the predicted offsets. In the experiments, we set M to M = 0.65|N(x̃)|, which can improve the fitting ability of the surface evolution network by driving the network to focus on the inaccurate predictions.

$L{\text{}}_{top\_k}\left(\theta \right)$

is used as the training loss function for the surface evolution network, and the stochastic gradient descent method is used to update the parameters of the network.

[0057] Step 5: surface evolution: deforming the initial surface points along the offsets to obtain the predicted surface, where the offsets present the prediction of the surface evolution network;

[0058] After training the surface evolution netwrok, input the points N(x̃) = {z.sub.i ∈ P|dist(z.sub.i,x̃) ≤ r} on initialization surface obtained in step 3 to the surface evolution network. To be noticed, the neighbor points N(x̃) of the surface point x̃ also contains other points near the surface. Thus, one can obtain the offsets of all the initial surface points without testing all the neighbor spheres. With the prediction offsets, the target position of the surface points is determined by adding the offsets to the coordinates of their initial position:

$u_i={\tilde{x}}_i+\gamma \left({\tilde{x}}_i\right),\forall {\tilde{x}}_i\in \tilde{S}$

where x̃.sub.i presents the coordinates of the position of the initial position for the points on surface, γ(x̃.sub.i) presents the offsets predicted by the surface evolution network, and S̃ is the set of initial surface points. For a three-dimensional image, the position of each pixel can be uniquely determined by the three-dimensional coordinates, and the offsets are also three-dimensional vectors whose values and signs represent the moving distance and the direction along three dimensions, respectively. Therefore, the location of the predicted surface points can be uniquely determined by the coordinates of each initialization point and the corresponding predicted offset.

[0059] The second network predicts the offsets of the surface points. The initial surface is evolved according to the offsets. Thus, the set of predicted surface points U = {u.sub.1, u.sub.2,..., u.sub.N} are obtained.

[0060] Step 6: surface reconstruction: reconstructing the 3D volumes from the set of predicted surface points set to obtain the final segmentation results.

[0061] After obtaining the evolved surface points set U = {u.sub.1,u.sub.2,...,u.sub.N}, surface reconstruction is required, since the final result should be a binary segmentation with the same size of the original image, rather than points set.

[0062] First, we use the Delaunay triangulation to build up the triangular mesh from the predicted boundary points set, and obtain the alpha shape mesh by removing the tetrahedrons with a circumradius larger than 5. Then we use the subdivide algorithm to voxelize the alpha shape mesh to obtain a closed surface. Finally, the flood fill algorithm is used to fill the holes in the closed surface and the final 3D segmentation volumes are obtained.

[0063] Dice Similarity Coefficient (DSC), Hausdorff distance (HD) and computational time are used to evaluate the performance of the presented model. All the experiments are done on NVIDIA’s GPU Titan RTX using Pytorch.

[0064] Dice Similarity Coefficient (DSC) measures the similarity between two sets defined as follows:

$DSC\left(\tilde{P},P\right)=\frac{2\left|\tilde{P}\cap P\right|}{\left|\tilde{P}\right|+\left|P\right|},$

where P presents the final binary segmentation results predicted by the model, and P̃ presents associated label. The larger the DSC is, the better the segmentation result is.

[0065] The Hausdorff distance (HD) measures the distance between two surfaces defined as follows:

$HD\left(\tilde{P},P\right)=\max \left\{\underset{x\in \partial \tilde{P}}{\max }\underset{y\in \partial P}{\min }{‖x-y‖}^2,\underset{y\in \partial P}{\max }\underset{x\in \partial \tilde{P}}{\min }{‖x-y‖}^2\right\}$

where

$\partial P$

presents the points set of the predicted organ surface, and

$\partial \tilde{P}$

presents the points set of the ground truth organ surface. The smaller the HD is, the closer the two surfaces is, and the better the prediction result is.

[0066] The results of related methods were compared on the MSD (Medical Segmentation Decathlon) spleen dataset, which contains 41 volumes. We randomly divide the dataset into two groups, one group containing 21 volumes used for training and the other one with 20 volumes used for testing.

TABLE-US-00001 Results on the MSD spleen dataset Methods 2D methods RSTN 3D methods VNet 2D surface evolution methods EBP 2D surface evolution mthods LSM The present invention DSC (%) 89.70 92.94 92.79 93.03 95.21 95%HD (mm) 5.12 3.97 2.23

[0067] The two-dimensional methods RSTN was proposed by [Qihang Yu, Lingxi Xie, Yan Wang, Yuyin Zhou, Elliot K. Fishman, and Alan L. Yuille. Recurrent saliency transformation network: Incorporating multi-stage visual cues for small organ segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8289, June 2018. ]. The three-dimensional methods V-net was proposed by [Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 Fourth International Conference on 3D Vision (3DV), October 2016.] The two-dimensional boundary-based method EBP [Tianwei Ni, Lingxi Xie, Huangjie Zheng, Elliot K Fishman, and Alan L Yuille. Elastic boundary projection for 3d medical image segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2109-2118, 2019.] and the other one LSM proposed by [Lihong Guo, Yueyun Liu, Yu Wang, Yuping Duan, and Xue-Cheng Tai. Learned snakes for 3D image segmentation. Signal Processing, 183:108013, June 2021.] are used for comparison. The presented model achieved the best segmentation accuracy with an overall 2% higher DSC than the others. By comparing the HD obtained by the three boundary-based methods, the presented model gave the smallest HD no matter before or after the voxelization, which means the presented model evolved surface closer to the target surface. FIG. 5 compares the label, EBP method, LSM method on two typical examples of the MSD spleen dataset. It can be seen from comparison that the segmentation of the presented invention is more accurately at the boundary, while the other two boundary based methods tend to over segment the spleen region.

[0068] Another public spleen dataset was also used, which contains 90 CT volumes with 43 from The Cancer Imaging Archive (TCIA) and 47 from the Beyond The Cranial Vault (BTCV) segmentation challenge. We randomly chose 45 volumes with 21 volumes from TCIA and 24 volumes from BTCV to train our DAS model, while the remaining 45 volumes were used for evaluation.

TABLE-US-00002 The testing results on the public spleen dataset Methods DSC (%) 95%HD (mm) Time (s) 2D surface evolution methods LSM 91.80% 7.84 213 The present invention 93.87% 4.69 76

[0069] As we can see from the Table 2, the presented model not only gives more than 2% higher DSC than LSM, but also saves more than half the computational time. Two segmentation results are displayed in FIG. 6 for a visual comparison. It is clearly shown that our model can improve the segmentation accuracy, while LSM tends to over segment the spleen region.

[0070] Whilst the foregoing description describes exemplary embodiments, it will be understood by those skilled in the art hat many variations of the embodiment can be made within the scope and spirit of the present invention.