Intestinal centerline extraction method based on prior feature learning and dual-direction guidance shortest path model

Abstract

The invention discloses an intestinal tract center line extraction method based on prior feature learning and a dual-direction guidance shortest path model, and relates to the technical field of tubular structure analysis in medical images. According to the method, the prior features are extracted by using the neural network, and the measurement map is constructed in combination with the shortest circuit model for solving the final center line. Firstly, a distance field, a dual direction field and a segmentation map are predicted through a multi-branch network. Then, centerline points are sampled from the output of the decoder branches using non-maximum suppression, and a Finler metric is constructed using the learned anatomical information to construct a shortest path model with direction guidance to calculate geodesic lines between the centerline points. And constructing a geodesic map by using the center line points and geodesic lines between the center line points. And finally, determining a final center line by solving a Hamiltonian path problem. By means of the method, the intestinal tract center line can be extracted more effectively and automatically, a more accurate, more complete and more robust center line extraction result is obtained, and the method has high clinical application value.

Description

Technical Field

[0001] This invention belongs to the field of medical image analysis technology, specifically relating to a method for extracting the intestinal centerline based on prior feature learning and a dual-direction guided shortest path model. Background Technology

[0002] Gastrointestinal diseases, such as inflammatory bowel disease (IBD), Crohn's disease, and colorectal cancer, are significant health problems worldwide. For example, colorectal cancer alone caused approximately 935,000 deaths in 2020. These diseases not only cause physical suffering but also severely impact patients' quality of life due to chronic pain, malnutrition, and other complications. Given that the gastrointestinal tract is located deep within the body and cannot be directly observed with the naked eye, advanced imaging techniques are crucial for diagnosis and monitoring. Traditional X-ray and ultrasound methods have limited effectiveness in treating complex gastrointestinal diseases. More precise imaging techniques, such as computed tomography (CT), magnetic resonance imaging (MRI), and endoscopy, are often required. These techniques can provide detailed images of internal structures, but interpretation requires specialized knowledge and tools.

[0003] With advancements in medical imaging technology, intelligent assisted analysis tools are playing an increasingly important role in diagnosing gastrointestinal diseases. Accurate segmentation and centerline extraction are fundamental to these tools, helping physicians understand the geometry of the gastrointestinal tract and perform quantitative analysis. For example… Figure 2 As shown, the extracted centerline aids in virtual navigation, helping surgeons plan the optimal surgical path preoperatively, thereby reducing risks and increasing success rates. Therefore, accurate segmentation results and centerlines not only improve diagnostic accuracy but also support the development of personalized treatment plans.

[0004] Manually extracting the intestinal centerline requires a high level of expertise, making the collection of sufficient data both expensive and challenging. Furthermore, the process is time-consuming and labor-intensive. Therefore, automated centerline extraction has attracted widespread attention, with traditional methods falling into three main categories: 1) Shortest Path Model: This method utilizes the shortest path algorithm from graph theory, treating the gastrointestinal tract as a graph where nodes represent pixels and edge weights are determined by pixel-wise metrics. While effective in finding paths from one endpoint to another, this method is sensitive to image noise and cannot efficiently handle bifurcation problems. 2) Generalized Potential Field Method: This technique simulates the potential field distribution in physics to guide centerline generation. It calculates the potential value at each point and proceeds along the gradient direction to find the centerline. Compared to distance fields, the main advantage of these functions is that they can produce satisfactory curves over a mid-range region of the target object where the distance field is constant. This is because they consider a larger boundary region, not just the distance to the nearest boundary. This also produces an averaging effect, making these algorithms less sensitive to boundary noise. While it performs well in handling complex shapes, its performance is highly dependent on initial parameter settings; improper configuration can lead to inaccurate results. Furthermore, it is computationally expensive. 3) The thinning method first uses a segmentation algorithm to obtain a binary mask of the gastrointestinal region, then applies morphological operations to gradually shrink the mask until a single-pixel-wide skeleton, i.e., the centerline, is obtained. This method is simple and direct, but may exhibit biases when handling uneven thickness variations or artifacts.

[0005] In recent years, deep learning methods have achieved tremendous success in various tasks, significantly surpassing traditional methods. Many researchers have designed deep learning-based algorithms for intestinal centerline extraction and segmentation. Some researchers have attempted to use neural networks as trackers, iteratively predicting the next step from a starting point and obtaining the entire centerline when tracking stops. However, due to the high complexity of the task, the accuracy of their extracted intestinal data is not satisfactory.

[0006] In summary, complete intestinal segmentation and centerline extraction remain challenging, mainly due to the following reasons:

[0007] 1) Complex morphology: The intestines are coiled and tortuous in the abdominal cavity, making it difficult to determine the sequential relationship between different intestinal segments and to design a measurement method that can well express the characteristics of the intestines.

[0008] 2) Intestinal overlap: The boundary between the outer wall of the intestine and the surrounding tissue is blurred, especially in the small intestine region. The segments of the intestine are closely spaced and difficult to distinguish, which can easily cause the central line to pass through the intestinal wall and produce a short circuit.

[0009] 3) Extremely long: As a tubular organ that can reach 6–8 meters in length, the intestine is the longest organ in the human body. During the extraction process, it is easy to miss some parts, that is, some intestinal segments are not completely extracted.

[0010] Recent research has attempted to learn the mapping from input to centerlines using deep neural networks trained on labeled datasets. While there is still significant room for improvement in the accuracy of their extracted centerlines and the sequentiality cannot be guaranteed, this approach demonstrates that the powerful representational capabilities of neural networks can be used to directly predict prior features. These prior features exhibit good interpretability. Summary of the Invention

[0011] Purpose of the invention: Traditional methods for extracting intestinal midline are difficult to design with appropriate metrics, leading to problems such as taking shortcuts and omissions. This invention can use prior-based metrics to solve these problems, enabling more efficient and automatic extraction of intestinal midline, resulting in more accurate, complete, and robust midline extraction results.

[0012] Technical Solution: This invention proposes a method for extracting the intestinal centerline based on prior feature learning and a dual-direction guided shortest path model. This method first uses a neural network to extract prior features, and then combines this with a shortest path model to construct a survey map to extract centerlines with complex shapes. It is highly interpretable and includes the following steps:

[0013] (a) Utilize a multi-task network to learn prior features with anatomical information, including geodesic distance field and dual-direction field (vector field), and extract centerline points;

[0014] (b) Construct a direction-guided shortest path model based on anatomical priors using the Finsler metric, convert the manually designed metric into a learning-based metric, and calculate the geodesics between each centerline point;

[0015] (c) Construct a geodesic distance map using centerline points and geodesics, and solve the Hamiltonian path problem to extract the final centerline, thus alleviating the problem of missing intestinal segments due to the excessive length of the intestine.

[0016] Furthermore, the multi-task network architecture proposed in step (a) is a Convolutional Neural Network (CNN) with residual connections between an encoder and decoder. The encoder consists of five layers, processing feature maps at different resolutions, with the number of channels increasing from 16 to 256. The encoder is used to extract general features, which are then fed into three independent task-specific decoder branches. Each branch is a four-layer decoder; each layer includes a 2×2 upsampled convolution, a summation operation with the relevant feature map via skip connections, and several 3×3 convolutional layers. The three decoders have the same structure, except that their final layer output channels differ: distance regression has 2 channels, dual vector field has 6 channels, and segmentation prediction has 2 channels.

[0017] Furthermore, in step (b), the prior features learned by the neural network... and We construct a new Finsler metric that incorporates prior geometric knowledge. First, for any point in the image space... It corresponds to a direction vector v θ =(v t +v n ) / ‖v t +v n ‖, where v t It is the tangent vector, v n It is a normal vector. A Finsler metric of the following form is defined on the image region Ω.

[0018]

[0019] Where ∝ represents the direct proportionality operator. It can be applied to any property belonging to... The inner product of a vector u is computed in the tangent space of Ω. However, the Finsler metric... Due to its high singularity, directly applying the fast traversal method would be difficult to compute the global minimum. Therefore, a family of directionally guided Finsler metrics dependent on the penalty parameter λ >> 1 is introduced, defined as follows:

[0020]

[0021] satisfy As λ→∞.

[0022] Furthermore, in order to make further use of more prior knowledge information, the predicted centerline distance field will be... As a function of velocity Used in this context. The learning-based Finsler metric can be defined as:

[0023]

[0024] It can be used To measure the length of curve γ:

[0025]

[0026]

[0027] This learning-based Finsler metric It is asymmetric, meaning that for most non-zero vectors u≠0, it usually satisfies:

[0028]

[0029] This asymmetry facilitates the construction of a directional centerline using a series of directional guide points. (Compared to the learning-based Finsler metric) Related, and the minimum action graph starting from the initial source point s It is the unique viscous solution to the following Eikonal partial differential equation:

[0030]

[0031] in yes The dual norm of . The geodesic distance between points x and s. It also relies on prior knowledge and the direction information of the shortest path. When λ is large enough, the differential equation (*) can be solved efficiently to obtain the minimum path.

[0032] Furthermore, in step (c), the extreme points are extracted from prior features. Construct a graph G = (V, E) as a set of nodes, where E is the set of edges. Between two nodes x... i and x j Create an edge e between them i,j Metrics can be used. The geodesic distance matrix is ​​calculated as follows:

[0033]

[0034] Node x i and x j The edge cost c between i,j Defined as

[0035]

[0036] Where ∈ is a small constant. γ * The shortest path found by the Fast Marching Method (FMM) satisfies...

[0037]

[0038] in The function cost(·) represents the total length of the connecting path. d(x) i ,x j ) represents x i and x j The Euclidean distance between them, δ is the distance threshold.

[0039] The beneficial effects of this invention are as follows: When faced with the task of intestinal centerline extraction, this invention uses prior-based metrics to address the difficulties in designing suitable metrics in traditional methods, which can lead to shortcuts and omissions. This allows for more efficient and automated extraction of the intestinal centerline, resulting in more accurate, complete, and robust centerline extraction results. It significantly reduces the workload of manual centerline annotation and has high clinical application value. Attached Figure Description

[0040] Figure 1 Here is an overview diagram of the method proposed in this invention: In the first stage, the network learns interpretable prior features (centerline distance field, dual direction field); then, based on the prior features output by the network, centerline nodes are extracted, and a shortest path model under Finsler metric incorporating direction information is constructed, followed by calculating the geodesic distance between each node; a geodesic map is constructed using centerline points and geodesics, and after specifying the start and end points, the Hamiltonian path problem is solved to find an optimal path that passes through all points in the map only once, and this path serves as the final centerline.

[0041] Figure 2 This is a diagram of the colon (green) and its center line (red). The center line has been expanded for clarity of visualization.

[0042] Figure 3 This is a comparison chart of centerline extraction results for various methods on the SICL dataset. Detailed Implementation

[0043] The objectives, technical solutions, and advantages of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] This invention proposes a method for extracting the intestinal centerline based on prior feature learning and a dual-direction guided shortest path model. This method first utilizes a neural network to extract prior features, and then combines this with a shortest path model to construct a survey map to extract centerlines with complex shapes, exhibiting high interpretability. The overall flowchart of this invention is shown below. Figure 1 As shown, it specifically includes the following content.

[0045] (a) The proposed multi-task network architecture is a Convolutional Neural Network (CNN) with residual connections between an encoder and decoder. The encoder consists of five layers, processing feature maps at different resolutions, with the number of channels increasing from 16 to 256. The encoder is used to extract general features, which are then fed into three independent task-specific decoder branches. Each branch is a four-layer decoder; each layer includes a 2×2 upsampled convolution, a summation operation with the relevant feature map via skip connections, and several 3×3 convolutional layers. The three decoders have the same structure, except that their final layer output channels differ: distance regression has 2 channels, dual vector field has 6 channels, and segmentation prediction has 2 channels.

[0046] (b) Prior features learned by neural networks and We construct a new Finsler metric that incorporates prior geometric knowledge. First, for any point in the image space... It corresponds to a direction vector v θ =(v t +v n ) / ‖v t +v n ‖, where v t It is the tangent vector, v n It is a normal vector. A Finsler metric of the following form is defined on the image region Ω.

[0047]

[0048] Where ∝ represents the direct proportionality operator. It can be applied to any property belonging to... The inner product of a vector u is computed in the tangent space of Ω. However, the Finsler metric... Due to its high singularity, directly applying the fast traversal method would be difficult to compute the global minimum. Therefore, a family of directionally guided Finsler metrics dependent on the penalty parameter λ >> 1 is introduced, defined as follows:

[0049]

[0050] satisfy As λ→∞.

[0051] Furthermore, in order to make further use of more prior knowledge information, the predicted centerline distance field will be... As a function of velocity Used in this context. The learning-based Finsler metric can be defined as:

[0052]

[0053] It can be used To measure the length of curve γ:

[0054]

[0055] This learning-based Finsler metric It is asymmetric, meaning that for most non-zero vectors u≠0, it usually satisfies:

[0056]

[0057] This asymmetry facilitates the construction of a directional centerline using a series of directional guide points. (Compared to the learning-based Finsler metric) Related, and the minimum action graph starting from the initial source point s It is the unique viscous solution to the following Eikonal partial differential equation:

[0058]

[0059] in yes The dual norm of . The geodesic distance between points x and s. It also relies on prior knowledge and the direction information of the shortest path. When λ is large enough, the differential equation (*) can be solved efficiently to obtain the minimum path.

[0060] (c) Extreme points extracted from prior features Construct a graph G = (V, E) as a set of nodes, where E is the set of edges. Between two nodes x... i and x j Create an edge e between them i,j Metrics can be used. The geodesic distance matrix is ​​calculated as follows:

[0061]

[0062] Node x i and x j The edge cost c between i,j Defined as

[0063]

[0064] Where ∈ is a small constant. γ * The shortest path found by the Fast Marching Method (FMM) satisfies...

[0065]

[0066] in The function cost(·) represents the total length of the connecting path. d(x) i ,x j ) represents x i and x j The Euclidean distance between them, δ is the distance threshold.

[0067] To more clearly illustrate the feasibility and superiority of this invention, seven different centerline extraction methods were compared in experiments on the SICL intestinal dataset: SwinUNETR V2, UNETR++, MP-grey, MP-mj, MP-grey-mask, MP-mj-mask, and the Learning-pMP model based on prior feature learning proposed in this invention.

[0068] Table 1 presents the quantitative evaluation results of seven centerline extraction methods on the small intestine dataset. MP-grey is the baseline model of this invention, and its index CL... quality CL Corr CL Comp and The scores are 0.0164, 0.1212, 0.0186, and 21.98, respectively. MP-mj uses enhanced intensity features, and its performance metrics are similar to MP-grey. For the MP-grey-mask and MP-mj-mask models, which introduce segmentation masks as prior knowledge constraints, the accuracy is improved, especially in CL. Corr The above method performs better. This indicates that the extracted centerline is closer to the true centerline. Among all methods, the Learning-pMP proposed in this invention performs best, with its CL... quality and The metrics are 0.7583 and 4.58, respectively. Compared to MP-mj-mask and UNETR++, CL... quality The accuracy of centerline extraction was significantly improved by approximately 0.62 and 0.27. This improvement is attributed to the invention's greater utilization of prior features extracted by the network, which provides stronger guidance compared to traditional filter-based low-level intensity enhancement methods, effectively mitigating challenges such as the "shortcut problem".

[0069] Table 1. Quantitative experimental results of centerline extraction on the intestinal dataset. Bold values ​​indicate the best results at statistical significance (p<0.05).

[0070]

[0071] Figure 3The visualization results of centerline extraction using seven methods are presented. The small intestine is the most complex intestinal organ, posing a significant challenge to centerline extraction. It has the longest total length, the widest coverage area, a tortuous shape, and its outer walls often overlap, making boundary distinction difficult. The visualization results show that the four MP methods perform poorly in this complex task, failing to fully extract the small intestine structure, mainly due to short-circuiting issues, resulting in impractical extraction results. Compared to MP-grey and MP-mj, MP-grey-mask and MP-mj-mask show slight improvements, but the gains are limited. This is because a significant characteristic of the small intestine is the overlapping of its outer walls, leading to blurred boundary edges in the segmentation mask. Therefore, the centerlines extracted by these methods easily penetrate the outer wall, causing short-circuiting problems. The centerlines obtained by refining the segmentation results of SwinUNETR V2 and UNETR++ contain many false positives, leading to low centerline accuracy (CL). Comp The values ​​are high, and some cluster-like structures also appear.

[0072] In comparison, the extraction results of the method of this invention are most consistent with the true centerline, achieving an almost complete extraction effect with good shape preservation. This advantage stems from the introduction of a centerline distance field and bidirectional fields to guide the direction of travel, thereby better capturing the edges and orientation of the small intestine. In addition, by introducing a set of centerline nodes and geodesic maps, the entire small intestine is divided into multiple local regions, effectively alleviating the problem of edge penalty failure caused by excessive geodesic distance when tracking very long tubular structures.

[0073] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A method for extracting the intestinal centerline based on prior feature learning and a dual-direction guided shortest path model, characterized in that, Includes the following steps: (a) Utilize a multi-task network to learn prior features with anatomical information, including segmentation maps, geodesic distance fields, and dual-direction fields, and extract centerline points; (b) Construct a direction-guided shortest path model based on anatomical priors using the Finsler metric, convert the manually designed metric into a learning-based metric, and calculate the geodesics between each centerline point; (c) Construct a geodesic distance graph using centerline points and geodesics, where the length of the geodesics is used as the weight of the edges, and solve the Hamiltonian path problem to extract the final centerline, thus alleviating the problem of missing intestinal segments caused by the excessive length of the intestine.

2. The intestinal centerline extraction method based on prior feature learning and a dual-direction guided shortest path model according to claim 1, characterized in that: The multi-task network in step (a) is a convolutional neural network with an encoder-decoder structure and residual connections. The encoder consists of five layers that process feature maps of different resolutions, with the number of channels increasing from 16 to 256. The encoder is used to extract general features and inputs these features into three independent task-specific decoder branches. Each branch is a four-layer decoder. Each layer includes a 2×2 upsampled convolution, a summation operation with the relevant feature map via skip connections, and several 3×3 convolutional layers. The three decoders have the same structure except that their final layer output channels are different: distance regression has 2 channels, dual vector field has 6 channels, and segmentation prediction has 2 channels.

3. The intestinal centerline extraction method based on prior feature learning and a dual-direction guided shortest path model according to claim 2, characterized in that: In step (b), the prior features learned by the neural network and Construct a new Fensler metric that integrates prior geometric knowledge; first, for any point in the image space... It corresponds to a direction vector v θ =(v t +v n ) / ‖v t +v n ‖, where v t It is the tangent vector, v n It is a normal vector; the Finsler metric is defined on the image region Ω in the following form. Where ∝ represents the direct proportionality operator; for any property belonging to The inner product of the vector u is computed in the tangent space of Ω; and a family of direction-guided Finsler metrics dependent on the penalty parameter λ >> 1 is introduced, defined as follows: satisfy As λ→∞. In addition, the predicted centerline distance field As a function of velocity Used; in this case, the learning-based Finsler metric is defined as: And under this metric, the shortest path connecting the two points, i.e., the geodesic, is calculated.

4. The intestinal centerline extraction method based on prior feature learning and a dual-direction guided shortest path model according to claim 3, characterized in that: In step (c), the extreme points are extracted from prior features. Construct a graph G = (V, E) as a set of nodes, where E is the set of edges; between two nodes x i and x j Create an edge e between them i,j Use metrics The geodesic distance matrix is ​​calculated as follows: Node x i and x j The edge cost c between i,j Defined as Where ∈ is a small constant; γ * The shortest path found by the fast traversal method satisfies in The function cost(·) is the total length of the connecting path; d(x i ,x j ) represents x i and x j The Euclidean distance between them, δ is the distance threshold.

Images