Artificial Intelligence-Based Method and System for Predicting the Difficulty of Rectal Cancer Surgery

Abstract

This invention provides an artificial intelligence-based method and system for predicting the difficulty of rectal cancer surgery. The method includes: acquiring high-resolution MRI images of the rectum; using a semantic analysis model to segment target regions of pelvic anatomical structures from the high-resolution MRI images of the rectum; extracting morphological features of the pelvic anatomical structures based on three-dimensional point cloud data determined from the target regions; constructing a graph structure model corresponding to the pelvic anatomical structures based on the three-dimensional point cloud data, and using a graph neural network to infer the graph structure model to obtain the graph features of the pelvic anatomical structures; and using a surgical difficulty prediction model to predict the baseline features, graph features, and morphological features of an individual patient to obtain the predicted difficulty of rectal cancer surgery and the surgical approach decision-making result. This invention can solve the problems of low prediction accuracy, poor stability, and low automation in the existing technology for predicting the difficulty of rectal cancer surgery.

Description

Technical Field

[0001] This invention relates to the field of medical technology, and in particular to a method and system for predicting the difficulty of rectal cancer surgery based on artificial intelligence. Background Technology

[0002] In recent years, with the deepening of the concept of precision surgery and the widespread application of laparoscopic surgery, the prediction of surgical difficulty and the optimization of surgical procedures for patients with mid-to-low rectal cancer have gradually become a hot topic in clinical research. Numerous studies have shown that factors such as gender, tumor location and size, pelvic diameter and angle, abdominal surgical history, and neoadjuvant therapy all have a significant impact on the difficulty of rectal cancer surgery.

[0003] Existing research analyzes four key factors influencing pelvic separation time: the sacrococcygeal diameter, the height of the sacrococcygeal arch, the ischial tuberosity diameter, and tumor size, classifying the difficulty of laparoscopic rectal cancer surgery into three levels: simple, moderate, and difficult. Furthermore, methods for assessing surgical difficulty based on indicators such as operative time, intraoperative blood loss, and conversion from laparoscopy to open surgery have been proposed, providing a quantitative reference for the difficulty level of the procedure. However, there are still some shortcomings in high-precision, fully automated prediction. Firstly, the accuracy and stability of surgical difficulty prediction are low. The difficulty of laparoscopic rectal cancer surgery is influenced by a combination of factors. Previous techniques relying on key intraoperative indicators (such as operative time and intraoperative blood loss) have significant limitations. On the one hand, these intraoperative indicators are affected by multiple factors and cannot fully reflect the impact of the patient's anatomical characteristics on the surgical difficulty; on the other hand, these indicators can only be used for subjective difficulty assessments and are difficult to accurately predict surgical difficulty. Secondly, the automation level of surgical difficulty prediction is low, resulting in insufficient practicality. Technical methods based on key measurement indicators or those based on bone structure analysis heavily rely on manual measurement and annotation of MRI images. This labor-intensive work is not only time-consuming and labor-intensive, but also prone to introducing human error. Summary of the Invention

[0004] This invention provides an artificial intelligence-based method and system for predicting the difficulty of rectal cancer surgery, which solves the problems of low prediction accuracy, poor stability and low degree of automation in the existing technology for predicting the difficulty of rectal cancer surgery.

[0005] This invention provides an artificial intelligence-based method for predicting the difficulty of rectal cancer surgery, comprising the following steps: Acquire high-resolution MRI images of the rectum; The semantic analysis model is invoked to segment the target region of pelvic anatomical structures from the high-resolution MRI image of the rectum; Based on the three-dimensional point cloud data determined from the target region, morphological features of the pelvic anatomical structures are extracted; Based on the three-dimensional point cloud data, a graph structure model corresponding to the pelvic anatomical structure is constructed, and a graph neural network is called to infer the graph structure model to obtain the graph features of the pelvic anatomical structure. The surgical difficulty prediction model is invoked to predict the baseline characteristics, graph features, and morphological features of an individual patient, thereby obtaining the predicted difficulty of rectal cancer surgery and the auxiliary decision-making results for surgical approach.

[0006] In some embodiments, after acquiring high-resolution MRI images of the rectum, the method further includes: The high-resolution MRI images of the rectum were standardized. The Gaussian filtering algorithm is used to filter the standardized high-resolution rectal MRI image, and the interpolation algorithm is used to resample the filtered high-resolution rectal MRI image to the same image resolution.

[0007] In some embodiments, the step of invoking a semantic analysis model to segment the target region of pelvic anatomical structures from the high-resolution MRI image of the rectum includes: The encoder in the semantic analysis model is invoked to extract multi-scale features from the high-resolution MRI image of the rectum. The decoder in the semantic analysis model is invoked to perform deconvolution processing on the multi-scale features to obtain the semantic segmentation results of the pelvic anatomical structure. The semantic segmentation results are subjected to morphological closing operations, and the processed semantic segmentation results are merged using a connected component analysis algorithm to form the target region of the pelvic anatomical structure.

[0008] In some embodiments, extracting morphological features of pelvic anatomical structures based on three-dimensional point cloud data determined from the target region includes: Extract the region of interest from the segmentation mask of the target region, and convert the pixels of the region of interest into three-dimensional point cloud data; Three-dimensional reconstruction is performed based on the three-dimensional point cloud data to obtain a triangular mesh representing the anatomical structure of the pelvis. The surface area, volume, and curvature distribution of the triangular mesh are calculated as morphological features of the pelvic anatomical structure.

[0009] In some embodiments, constructing a graphical structural model corresponding to the pelvic anatomical structure based on the three-dimensional point cloud data includes: The three-dimensional point cloud data corresponding to the pelvic anatomical structure is used as nodes, the positional relationship between the nodes is used as the edges of the nodes, and the geometric features of the nodes are used as the node attributes of the nodes. The geometric features include at least one of surface area, volume, and curvature. Based on the nodes, edges, and node attributes, a graph structure model corresponding to the pelvic anatomical structure is constructed.

[0010] In some embodiments, the step of calling the surgical difficulty prediction model to predict the patient's baseline characteristics, the graphical features, and the morphological features to obtain the difficulty prediction results for rectal cancer surgery and the surgical approach auxiliary decision results includes: Obtain the baseline characteristics of individual patients; The baseline features, the graph features, and the morphological features are concatenated to obtain the fused features; The surgical difficulty prediction model is invoked to predict the fused features, thereby obtaining the difficulty prediction result of rectal cancer surgery, and the surgical approach is assisted in decision-making based on the difficulty prediction result. The methods for assisting surgical approach decision-making include: When the difficulty prediction result is greater than or equal to the preset threshold, the surgical approach is selected as the transanal approach; when the difficulty prediction result is less than the preset threshold, the surgical approach is selected as the transabdominal approach.

[0011] In some embodiments, obtaining the baseline characteristics of an individual patient includes: Determine the individual patient's physical numerical characteristics and treatment category characteristics; The normalized body numerical features and the embedded coded treatment category features are used as the baseline features of individual patients.

[0012] This invention also provides an artificial intelligence-based system for predicting the difficulty of rectal cancer surgery, comprising the following steps: The acquisition module is used to acquire high-resolution MRI images of the rectum. The segmentation module is used to call a semantic analysis model to segment the target region of pelvic anatomical structures from the high-resolution MRI image of the rectum; The extraction module is used to extract morphological features of pelvic anatomical structures based on three-dimensional point cloud data determined from the target region. The inference module is used to construct a graph structure model corresponding to the pelvic anatomical structure based on the three-dimensional point cloud data, and call the graph neural network to infer the graph structure model to obtain the graph features of the pelvic anatomical structure. The prediction module is used to call the surgical difficulty prediction model to predict the baseline features, graph features, and morphological features of an individual patient, so as to obtain the difficulty prediction results of rectal cancer surgery and the surgical approach auxiliary decision results.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the artificial intelligence-based method for predicting the difficulty of rectal cancer surgery as described above.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the artificial intelligence-based method for predicting the difficulty of rectal cancer surgery as described above.

[0015] This invention provides an AI-based method for predicting the surgical difficulty of rectal cancer. Targeting high-resolution rectal MRI images, it utilizes an AI-based semantic analysis model, graph neural network, and surgical difficulty prediction model to achieve accurate and stable predictions of the surgical approach. This helps in decision-making regarding surgical approach, overcoming the limitations of existing technologies that focus only on a single anatomical structure or feature. Furthermore, during the prediction process, semantic segmentation of pelvic anatomy is seamlessly integrated with surgical difficulty prediction, achieving full automation from high-resolution rectal MRI image data to clinical decision support. Attached Figure Description

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

[0017] Figure 1 This is a flowchart illustrating the artificial intelligence-based method for predicting the difficulty of rectal cancer surgery provided by this invention.

[0018] Figure 2 This is a schematic diagram of the structure of the artificial intelligence-based rectal cancer surgery difficulty prediction system provided by the present invention.

[0019] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] The artificial intelligence-based method for predicting the difficulty of rectal cancer surgery according to the present invention will be described below with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating the artificial intelligence-based method for predicting the difficulty of rectal cancer surgery provided by this invention. Figure 1 As shown, the method includes the following steps 101 to 105.

[0022] Step 101: Obtain high-resolution MRI images of the rectum for rectal cancer.

[0023] Here, during the rectal cancer examination process, MRI images of individual pelvic anatomical structures are collected using medical equipment as basic data for assessing the difficulty of rectal cancer surgery.

[0024] In some embodiments, after acquiring high-resolution MRI images of the rectum for rectal cancer, preprocessing is required. Specifically, the high-resolution MRI images of the rectum are standardized. During standardization, the mean and standard deviation of the pixels in the high-resolution MRI images are calculated, and then normalized. The formula is as follows: (1) Where I represents a high-resolution MRI image of the rectum. This represents the mean value of the corresponding pixel in a high-resolution MRI image of the rectum. This represents the standard deviation of a corresponding pixel in a high-resolution rectal MRI image, used to adjust pixel values ​​in high-resolution rectal MRI images to a zero-mean unit variance distribution. This represents a high-resolution MRI image of the rectum after standardization.

[0025] In addition, histogram equalization can be used to enhance the contrast of high-resolution MRI images of the rectum before standardization.

[0026] Furthermore, the Gaussian filtering algorithm is invoked to filter the standardized high-resolution rectal MRI image, and the interpolation algorithm is invoked to resample the filtered high-resolution rectal MRI image to the same image resolution.

[0027] Here, the Gaussian filtering algorithm can use a convolutional network based on a Gaussian kernel. The formula for Gaussian filtering is as follows: (2) in, This represents a high-resolution MRI image of the rectum after Gaussian filtering. Represents the Gaussian kernel function. This represents the convolution operation.

[0028] Regarding image resolution, resampling is performed using an interpolation algorithm to adjust the image size of the high-resolution rectal MRI images, ensuring all high-resolution rectal MRI images are processed to the same image resolution, thus completing image preprocessing. The interpolation algorithm can be a bicubic interpolation algorithm, and the interpolation process is expressed as follows: (3) in, This represents a preprocessed high-resolution MRI image of the rectum. These represent the height and width of the resampled high-resolution MRI image of the rectum, respectively. This represents the computation function for resampling algorithms, such as bicubic interpolation.

[0029] In this embodiment of the invention, the high-resolution rectal MRI images are processed by standardization, Gaussian filtering, and resampling, which can effectively remove high-frequency noise in the images, making the quality of the high-resolution rectal MRI images consistent, ensuring the standardization and robustness of the input data, and providing high-quality data for subsequent semantic analysis model segmentation.

[0030] Step 102: Use the semantic analysis model to segment the target region of pelvic anatomical structures from the high-resolution MRI image of the rectum.

[0031] Here, semantic segmentation of pelvic anatomical structures is achieved using preprocessed high-resolution rectal MRI images. To address the complexity of pelvic anatomy and the diversity of input image parameters, the semantic analysis model is designed with a dynamic adaptive architecture, which can be a neural network model used for image semantic segmentation. For example, the U-Net++ model has three downsampling levels. An adaptive convolutional module is added to the encoder, which dynamically adjusts the size of the convolutional kernel according to the image resolution of the input high-resolution rectal MRI image, for example, from 3×3×3 to 5×5×5. The decoder uses transposed convolution and a skip connection structure to ensure that feature details are not lost. The model loss function can use DiceLoss.

[0032] The semantic analysis model can segment the image region of pelvic anatomical structures in high-resolution MRI images of the rectum and remove irrelevant background areas.

[0033] In some embodiments, the semantic analysis model can be invoked to segment the target region of pelvic anatomical structures from high-resolution rectal MRI images. This can be achieved in the following ways, which are described in detail below.

[0034] The encoder in the semantic analysis model is invoked to extract multi-scale features from high-resolution MRI images of the rectum. Here, the semantic analysis model consists of a decoder and an encoder. The encoder is composed of a multi-layer convolutional network and is used to extract semantic feature maps of high-resolution rectal MRI images at multiple scales, as shown below: (4) Where F represents the extracted multi-scale features. This represents the encoder's processing function. This represents a preprocessed high-resolution MRI image of the rectum.

[0035] Furthermore, the encoder employs an Adaptive Convolution Module (ACM), where the kernel k and the number of channels c in each convolutional network layer can be dynamically adjusted according to the image resolution of the preprocessed high-resolution rectal MRI image, achieving adaptation, as shown below: (5) in, This indicates the image resolution of the preprocessed high-resolution MRI image of the rectum. This represents the image resolution contrast function. This represents a dynamic configuration function that adjusts the convolution kernel k and the number of channels c.

[0036] Furthermore, the decoder in the semantic analysis model is invoked to perform deconvolution processing on the multi-scale features, obtaining the semantic segmentation result of the pelvic anatomy structure. The semantic feature maps at each scale output by the encoder are input to the decoder. The decoder consists of a corresponding deconvolution network, which can progressively restore the image resolution from the semantic feature maps through deconvolution calculations, generating the semantic segmentation result, i.e., the semantic segmentation map of the pelvic anatomy structure. The decoding process is represented as follows: (6) Where S represents the semantic segmentation result of the pelvic anatomical structure. This represents the decoder's processing function.

[0037] Here, since the decoder performs multiple deconvolutions to adjust the image resolution of the semantic feature map, semantic information may be lost. Therefore, the skip connection module designed in the decoder can integrate the shallow semantic features with high image resolution into the deep semantic features with low image resolution, ensuring that the semantic feature details of the segmentation result are not lost and that it has a complete semantic expression.

[0038] Morphological closing operations are performed on the semantic segmentation results, and the processed semantic segmentation results are merged using a connected component analysis algorithm to form the target region of the pelvic anatomical structure.

[0039] Here, considering the potential artifacts or discontinuous segmentation regions in the semantic segmentation results, further refinement is needed. First, morphological closing operations are performed on the semantic segmentation results to eliminate artifacts as much as possible. During the morphological closing operation, structuring elements are used to perform dilation and erosion operations on the semantic segmentation graph, as shown below: (7) Where S represents the semantic segmentation result output by the decoder in the semantic analysis model, and B represents the structuring element used in the morphological closing operation. This indicates an expansion operation. This indicates a corrosion operation.

[0040] Furthermore, to integrate potentially disconnected regions in the semantic segmentation results, this embodiment of the invention uses a connected component analysis algorithm to merge the semantic segmentation results after morphological closing operations, that is, to merge these disconnected segmentation regions together, as shown below: (8) in, This represents the semantic segmentation result after merging, i.e., the target region of the pelvic anatomical structure. This represents the processing function of the connected component analysis algorithm.

[0041] This invention, through a semantic analysis model, achieves precise identification and segmentation of pelvic anatomical structures in high-resolution rectal MRI images by encoding and decoding. It can better distinguish between rigid structures (bones) and flexible structures (organs, tumors, etc.) within the pelvic anatomy, laying a reliable anatomical foundation for predicting the difficulty of subsequent surgeries. Furthermore, morphological closing operations and connected component analysis are performed on the semantic segmentation results to ensure smoothness and continuity, thereby improving segmentation accuracy.

[0042] Step 103: Extract the morphological features of the pelvic anatomical structures based on the three-dimensional point cloud data determined from the target region.

[0043] After determining the semantic segmentation map of the pelvic anatomical structure through step 102, it is further necessary to extract the morphological features of the pelvic anatomical structure from it, which will be used as a criterion for predicting surgical difficulty. Here, considering that the pelvic anatomical structure includes organs and tissues such as the left and right iliac bones, sacrum, rectal mesentery, rectum, and tumors, it has complex spatial distribution characteristics that can directly affect surgical planning. Two-dimensional images alone cannot accurately represent these spatial distribution characteristics. Therefore, in this embodiment of the invention, a three-dimensional modeling method is used to convert the pixels in the target area of ​​the pelvic anatomical structure into a three-dimensional point cloud. Then, a three-dimensional mesh model of the pelvic anatomical structure is constructed through three-dimensional reconstruction to clearly express the spatial distribution characteristics of each tissue and organ in the pelvic anatomical structure. Through geometric operations on the mesh model, the morphological features of the pelvic anatomical structure are extracted.

[0044] In some embodiments, the extraction of morphological features of pelvic anatomical structures based on three-dimensional point cloud data determined from the target region can be achieved in the following ways, as detailed below.

[0045] First, the region of interest (ROI) is extracted from the segmentation mask of the target region, and the pixels of the ROI are converted into 3D point cloud data. Since there may be some irrelevant background areas in the target region, the ROI is extracted from the segmentation mask of the target region; the ROI represents the area containing the various tissues and organs of the pelvic anatomical structure. Then, for each pixel in the ROI, it is converted into 3D point cloud data. This step can be achieved using 3D point cloud technology, or by extracting depth information from the ROI, obtaining the depth value of each pixel, and then calculating the 3D coordinates of each pixel, thus forming the 3D point cloud data, as shown below: (9) Where P represents the 3D point cloud data of the region of interest. Let N represent the i-th 3D point cloud data, and let N represent the total number of 3D point cloud data. They represent The three-dimensional coordinates.

[0046] Next, 3D reconstruction is performed based on the 3D point cloud data to obtain a triangular mesh representing the pelvic anatomical structure. Here, the Marching Cubes algorithm can be used to reconstruct the 3D point cloud data to build the surface mesh of the pelvic anatomical structure, and then Poisson surface reconstruction is used to optimize the point cloud density of the surface mesh. During the construction process, the surface network can be triangulated to finely represent the surface information of the mesh model.

[0047] Furthermore, the surface area, volume, and curvature distribution of the triangular mesh were calculated as morphological features of the pelvic anatomical structure.

[0048] The formula for calculating the surface area of ​​a triangular mesh is as follows: (10) Where T represents the set of triangular faces of the triangular mesh, and t represents the t-th triangular face. Let represent the area of ​​the t-th triangular facet.

[0049] The volume of a triangular mesh can be calculated based on the surface integration of the point cloud, for example, using an integral approximation method. The curvature distribution is used to describe the 3D point cloud data within the triangular mesh. The local geometric complexity characteristics are calculated using the following formula: (11) in, Representing 3D point cloud data curvature distribution, Representing 3D point cloud data Neighbor point cloud collection, Representing 3D point cloud data Neighbor Dianyun Indicates calculation and Euclidean distance.

[0050] The surface area, volume, and curvature distribution calculated here represent the actual surface area, volume, and curvature distribution of each organ and tissue structure in the pelvic anatomy, and are denoted as M as morphological characteristics.

[0051] In this embodiment of the invention, by calculating the surface area, volume, and curvature distribution as morphological features of the pelvic anatomical structure, the spatial distribution characteristics of the pelvic anatomical structure can be vividly characterized, laying a physical basis for predicting the difficulty of subsequent surgeries.

[0052] Step 104: Construct a graph structure model corresponding to the pelvic anatomical structure based on the 3D point cloud data, and call the graph neural network to infer the graph structure model to obtain the graph features of the pelvic anatomical structure.

[0053] Since the spatial relationships of various organs and tissues within the pelvic anatomical structure also have a significant impact on the difficulty of surgery, such as the distance between the sacrum and rectum, and the contact area between the tumor and the rectal mesentery, this embodiment of the invention also performs topological relationship modeling of the pelvic anatomical structure. Specifically, this can involve constructing a graph structure model and then calling a graph neural network algorithm for reasoning, using these topological relationships to capture and understand the synergistic effects between various organs and tissues within the pelvic anatomical structure.

[0054] In some embodiments, constructing a graphical model of the pelvic anatomical structure based on 3D point cloud data can be achieved in the following ways, as detailed below.

[0055] The three-dimensional point cloud data corresponding to the pelvic anatomical structure is used as nodes, the positional relationship between nodes is used as the edges of nodes, and the geometric features of nodes are used as node attributes. The geometric features include at least one of surface area, volume, and curvature.

[0056] This section constructs a graph structure model, first defining nodes by using the 3D point cloud data corresponding to anatomical tissues in the pelvic anatomy as nodes. These anatomical tissues can include rectal tumors, the mesentery, the iliac bone (left / right), and the sacrum. Thus, rectal tumors, the mesentery, the iliac bone (left / right), and the sacrum can be defined as five nodes. The positional relationships between nodes are used as edges, specifically Euclidean distances, such as the distance between the sacrum and the mesentery; this Euclidean distance is the edge weight. Furthermore, feature embedding is performed on the nodes, using their geometric features as node attributes. For example, the geometric features corresponding to a tumor are embedded as attributes representing the tumor node. Geometric features include at least one of surface area, volume, and curvature, such as the surface area, volume, and curvature of a tumor. The feature dimension of the geometric features is 128.

[0057] Based on nodes, edges, and node attributes, a graph structure model corresponding to the pelvic anatomical structure is constructed. Nodes are connected by edges, and node attributes are embedded, thereby transforming the 3D point cloud data of the pelvic anatomical structure into a graph structure model, represented as follows: (12) Where G represents the graph structure model, This represents a set of nodes, corresponding to the various organs and tissues within the pelvic anatomical structure. This represents the set of edges, which in turn represents the structural relationships between nodes.

[0058] In this embodiment of the invention, by converting the three-dimensional point cloud data of the pelvic anatomical structure into a graph structure model, the topological relationships between the organs and tissues can be expressed more accurately, laying the foundation for subsequent capture of the synergistic relationships between the organs and tissues.

[0059] Furthermore, the graph neural network is invoked to infer the graph structure model, obtaining the graph features of the pelvic anatomical structure. During the inference process, the graph neural network updates the node features of each node in the graph structure model, generating new node features. The formula for updating node features is as follows: (13) in, This represents the node features of node i before the update of the k-th graph convolutional layer in the graph neural network. This represents the node features of node i after being updated in the k-th graph convolutional layer of the graph neural network. , These represent the weights and biases of the k-th graph convolutional layer in the graph neural network, respectively. Let j represent the set of the i-th node's neighboring nodes, and j represent the neighboring nodes of node i. This represents the edge weight between node i and node j. This represents the activation function of the k-th graph convolutional layer in a graph neural network, such as ReLU.

[0060] The graphical features of the pelvic anatomical structure are extracted by performing a pooling operation on the node features obtained through inference, and are represented as follows: (14) in, The computation function representing the pooling operation. These represent the node features obtained from graph neural network inference. H represents the graph features of the pelvic anatomy, used to characterize and understand the overall spatial relationships between the organs and tissues in the pelvic anatomy.

[0061] In this embodiment of the invention, a structural model of pelvic anatomical tissue is constructed, and then a graph neural network is used to infer the spatial topological relationships between the structures of various organs and tissues, providing an effective spatial information basis for predicting the difficulty of surgery.

[0062] Step 105: Call the surgical difficulty prediction model to predict the baseline features, graph features and morphological features of the individual patient to obtain the predicted difficulty of rectal cancer surgery and the auxiliary decision-making results for surgical approach.

[0063] Here, this embodiment of the invention also considers that the prediction of surgical difficulty is related to the individual patient, and that objective clinical factors of the individual patient can also affect the surgical difficulty. Therefore, individual patient factors are also included in the characteristic factors for predicting surgical difficulty. The individual patient's objective clinical factors are represented by baseline characteristics, specifically including physical numerical characteristics and treatment category characteristics. Physical numerical characteristics include height, weight, and body mass index; treatment category characteristics include the individual patient's gender and whether the individual patient is willing to undergo adjuvant therapy for rectal cancer. These are all factors that influence the difficulty of rectal cancer surgery.

[0064] Based on the above factors, these baseline features also need to be processed when predicting surgical difficulty, and then merged with the morphological features M and graph features H extracted in the above steps to jointly achieve the prediction of surgical difficulty.

[0065] In some embodiments, the surgical difficulty prediction model is invoked to predict the baseline features, graph features, and morphological features of an individual patient to obtain the predicted difficulty of rectal cancer surgery. This can be achieved in the following ways, which are explained in detail below.

[0066] First, obtain the baseline characteristics of each individual patient.

[0067] Specifically, the first step is to determine the patient's individual physical characteristics and treatment category. Physical characteristics include height, weight, and body mass index (BMI), calculated based on height and weight. Treatment category characteristics include the patient's gender and whether the patient is willing to undergo adjuvant therapy for rectal cancer via surgery, which can be determined clinically.

[0068] Furthermore, the normalized body numerical features and the embedded coded treatment category features are used as the baseline features of the individual patient, denoted as B.

[0069] Body numerical features are numerical features. Different body numerical features have different numerical ranges and should be processed to the same scale. Therefore, normalization is performed, and the result is expressed as follows: (15) Where x represents the body's numerical characteristics. This represents the normalized numerical characteristics of the body. , These represent the minimum and maximum values ​​of the body's numerical characteristics, respectively.

[0070] Treatment category features are categorical features with a finite number of values. They are generally converted into continuous vector representations through embedding encoding. (16) in, This represents the embedding weight matrix used in the embedding encoding, where y represents the treatment category feature. This represents the treatment category feature after embedding and encoding.

[0071] Therefore, the normalized body numerical features and the embedded coded treatment category features serve as the clinical characteristics of individual patients.

[0072] In this embodiment of the invention, the baseline characteristics of individual patients are obtained to characterize the objective clinical factors of individual patients, thereby enhancing the rationality of surgical difficulty prediction, enabling the surgical difficulty prediction results to provide more accurate clinical decision support, and improving the clinical application value of surgical difficulty prediction results.

[0073] Finally, the procedure for predicting surgical difficulty is performed. To achieve a comprehensive assessment of surgical difficulty, this will be done as follows: Baseline feature B, graph feature H, and morphological feature M are concatenated to obtain the fused feature, denoted as Q, which is represented as follows: (17) Furthermore, the surgical difficulty prediction model is invoked to process the fused features, yielding the predicted difficulty of rectal cancer surgery. Here, by fusing baseline features B, graph features H, and morphological features M, the fused features can capture the synergistic effect among anatomical features, spatial topological information, and clinical baseline features.

[0074] The surgical difficulty prediction model can be a pre-trained fully connected neural network capable of predicting the fused features of the input, represented as: (18) in, This represents the activation function of the surgical difficulty prediction model, such as the sigmoid function or the softmax function. , These represent the weight matrix and bias term of the surgical difficulty prediction model, respectively. This indicates the predicted difficulty of rectal cancer surgery. Specifically, it can be an evaluation score of the difficulty of rectal cancer surgery, with a value ranging from 0 to 1.

[0075] Furthermore, the surgical approach is assisted in decision-making based on the difficulty prediction results. The methods for assisting in decision-making for surgical approach include: when the difficulty prediction result is greater than or equal to a preset threshold, the surgical approach is selected as the transanal approach; when the difficulty prediction result is less than the preset threshold, the surgical approach is selected as the transabdominal approach.

[0076] Here, the surgical approach for rectal cancer surgery includes both anal and abdominal approaches, which must be selected based on the predicted surgical difficulty. A preset threshold is used, for example, 0.5. However, other values, such as 0.6, can be set depending on the specific circumstances; this embodiment of the invention does not impose limitations on this. When the predicted difficulty of rectal cancer surgery is greater than or equal to the preset threshold of 0.5, it indicates a high level of difficulty and risk. To ensure a high success rate, the anal approach is chosen. Conversely, when the predicted difficulty is less than or equal to the preset threshold of 0.5, it indicates a lower level of difficulty and risk, and the abdominal approach can be selected. Thus, by determining the surgical difficulty based on the predicted results, an accurate decision is made regarding the surgical approach for rectal cancer.

[0077] This invention, through the fusion of baseline features, graphical features, and morphological features, achieves multi-feature fusion analysis of high-resolution MRI images of the rectum. This accurately captures rigid anatomical features, flexible anatomical features, and patient baseline information relevant to rectal cancer surgery, improving the model's predictive performance and stability. The final surgical difficulty prediction results provide a clear and intuitive understanding of the degree of difficulty in rectal cancer surgery, offering effective support for subsequent surgical planning decisions and risk assessment.

[0078] In summary, this invention, targeting high-resolution rectal MRI images for rectal cancer, achieves accurate and stable prediction of surgical difficulty through an artificial intelligence semantic analysis model, graph neural network, and surgical difficulty prediction model. The prediction process first extracts the target region of the pelvic anatomical structures, then acquires morphological features, graph features, and baseline features to accurately capture the morphological characteristics of the anatomical structures, the spatial topological relationships between anatomical structures, and clinical features. This not only eliminates reliance on manual annotation and measurement, significantly reducing errors caused by human intervention, but also greatly improves the accuracy, stability, and consistency of the prediction. It also overcomes the limitations of existing technologies that focus only on a single anatomical structure or feature. Furthermore, the semantic segmentation of the pelvic anatomical structures is seamlessly integrated with the surgical difficulty prediction process, achieving full automation from high-resolution rectal MRI image data to clinical decision support. This simplifies the prediction steps and reduces time costs. Finally, the fusion analysis of the prediction part enhances the clinical applicability and reliability of the model, providing efficient support for surgical planning and risk assessment. This technological breakthrough, achieving accurate and stable prediction results and full automation, significantly enhances clinical application value and promotes the widespread adoption of surgical difficulty assessment technology.

[0079] The artificial intelligence-based rectal cancer surgery difficulty prediction system provided by the present invention will be described below. The artificial intelligence-based rectal cancer surgery difficulty prediction system described below can be referred to in correspondence with the artificial intelligence-based rectal cancer surgery difficulty prediction method described above.

[0080] like Figure 2 As shown, the artificial intelligence-based rectal cancer surgery difficulty prediction system includes: acquisition module 201, segmentation module 202, extraction module 203, reasoning module 204, and prediction module 205.

[0081] Specifically, the acquisition module 201 is used to acquire high-resolution MRI images of the rectum; the segmentation module 202 is used to call a semantic analysis model to segment the target region of the pelvic anatomical structure from the high-resolution MRI images of the rectum; the extraction module 203 is used to extract the morphological features of the pelvic anatomical structure based on the three-dimensional point cloud data determined from the target region; the reasoning module 204 is used to construct a graph structure model corresponding to the pelvic anatomical structure based on the three-dimensional point cloud data, and call a graph neural network to reason about the graph structure model to obtain the graph features of the pelvic anatomical structure; the prediction module 205 is used to call a surgical difficulty prediction model to predict the baseline features of the individual patient, the graph features, and the morphological features to obtain the difficulty prediction result of rectal cancer surgery and the surgical approach auxiliary decision result.

[0082] In some embodiments, the acquisition module 201 is further configured to standardize the high-resolution rectal MRI image; call a Gaussian filtering algorithm to filter the standardized high-resolution rectal MRI image; and call an interpolation algorithm to resample the filtered high-resolution rectal MRI image to the same image resolution.

[0083] It should be noted that the beneficial effects of the AI-based rectal cancer surgery difficulty prediction system described above correspond to each other, so the beneficial effects of the AI-based rectal cancer surgery difficulty prediction system will not be elaborated here.

[0084] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3 As shown, the electronic device may include a processor 310, a communications interface 320, a memory 330, and a communication bus 340, wherein the processor 310, communications interface 320, and memory 330 communicate with each other via the communication bus 340. The processor 310 can call logical instructions in the memory 330 to execute an artificial intelligence-based method for predicting the difficulty of rectal cancer surgery. This method includes: acquiring high-resolution MRI images of the rectum; using a semantic analysis model to segment a target region of pelvic anatomical structures from the high-resolution MRI images of the rectum; extracting morphological features of the pelvic anatomical structures based on three-dimensional point cloud data determined from the target region; constructing a graph structure model corresponding to the pelvic anatomical structures based on the three-dimensional point cloud data, and using a graph neural network to infer the graph structure model to obtain graph features of the pelvic anatomical structures; and using a surgical difficulty prediction model to predict the patient's baseline features, the graph features, and the morphological features to obtain a predicted result for the difficulty of rectal cancer surgery and a surgical approach decision-making result.

[0085] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0086] On the other hand, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements an artificial intelligence-based method for predicting the difficulty of rectal cancer surgery provided by the methods described above. This method includes: acquiring high-resolution MRI images of the rectum; invoking a semantic analysis model to segment a target region of pelvic anatomical structures from the high-resolution MRI images of the rectum; extracting morphological features of the pelvic anatomical structures based on three-dimensional point cloud data determined from the target region; constructing a graph structure model corresponding to the pelvic anatomical structures based on the three-dimensional point cloud data, and inferring from the graph structure model using a graph neural network to obtain graph features of the pelvic anatomical structures; and invoking a surgical difficulty prediction model to predict the baseline features of an individual patient, the graph features, and the morphological features to obtain a predicted result for the difficulty of rectal cancer surgery and a decision-making result for the surgical approach.

[0087] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0088] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the surgical difficulty of rectal cancer based on artificial intelligence, characterized in that, include: Acquire high-resolution MRI images of the rectum; The semantic analysis model is invoked to segment the target region of pelvic anatomical structures from the high-resolution MRI image of the rectum; Based on the three-dimensional point cloud data determined from the target region, morphological features of the pelvic anatomical structures are extracted; Based on the three-dimensional point cloud data, a graph structure model corresponding to the pelvic anatomical structure is constructed, and a graph neural network is called to infer the graph structure model to obtain the graph features of the pelvic anatomical structure. The surgical difficulty prediction model is invoked to predict the baseline characteristics, graph features, and morphological features of an individual patient, thereby obtaining the predicted difficulty of rectal cancer surgery and the auxiliary decision-making results for surgical approach.

2. The method for predicting the surgical difficulty of rectal cancer based on artificial intelligence according to claim 1, characterized in that, After acquiring high-resolution MRI images of the rectum, the method further includes: The high-resolution MRI images of the rectum were standardized. The Gaussian filtering algorithm is used to filter the standardized high-resolution rectal MRI image, and the interpolation algorithm is used to resample the filtered high-resolution rectal MRI image to the same image resolution.

3. The method for predicting the surgical difficulty of rectal cancer based on artificial intelligence according to claim 1, characterized in that, The invocation of the semantic analysis model to segment the target region of pelvic anatomical structures from the high-resolution MRI image of the rectum includes: The encoder in the semantic analysis model is invoked to extract multi-scale features from the high-resolution MRI image of the rectum. The decoder in the semantic analysis model is invoked to perform deconvolution processing on the multi-scale features to obtain the semantic segmentation results of the pelvic anatomical structure. The semantic segmentation results are subjected to morphological closing operations, and the processed semantic segmentation results are merged using a connected component analysis algorithm to form the target region of the pelvic anatomical structure.

4. The method for predicting the surgical difficulty of rectal cancer based on artificial intelligence according to claim 1, characterized in that, The step of extracting morphological features of pelvic anatomical structures based on three-dimensional point cloud data determined from the target region includes: Extract the region of interest from the segmentation mask of the target region, and convert the pixels of the region of interest into three-dimensional point cloud data; Three-dimensional reconstruction is performed based on the three-dimensional point cloud data to obtain a triangular mesh representing the anatomical structure of the pelvis. The surface area, volume, and curvature distribution of the triangular mesh are calculated as morphological features of the pelvic anatomical structure.

5. The method for predicting the surgical difficulty of rectal cancer based on artificial intelligence according to claim 1, characterized in that, The construction of the graphical structural model corresponding to the pelvic anatomical structure based on the three-dimensional point cloud data includes: The three-dimensional point cloud data corresponding to the pelvic anatomical structure is used as nodes, the positional relationship between the nodes is used as the edges of the nodes, and the geometric features of the nodes are used as the node attributes of the nodes. The geometric features include at least one of surface area, volume, and curvature. Based on the nodes, edges, and node attributes, a graph structure model corresponding to the pelvic anatomical structure is constructed.

6. The method for predicting the surgical difficulty of rectal cancer based on artificial intelligence according to claim 1, characterized in that, The surgical difficulty prediction model is invoked to predict the patient's baseline characteristics, graphical features, and morphological features, resulting in a prediction of the difficulty of rectal cancer surgery and an auxiliary decision-making result for the surgical approach, including: Obtain the baseline characteristics of individual patients; The baseline features, the graph features, and the morphological features are concatenated to obtain the fused features; The surgical difficulty prediction model is invoked to predict the fused features, thereby obtaining the difficulty prediction result of rectal cancer surgery, and the surgical approach is assisted in decision-making based on the difficulty prediction result. The methods for assisting surgical approach decision-making include: When the difficulty prediction result is greater than or equal to the preset threshold, the surgical approach is selected as the transanal approach; when the difficulty prediction result is less than the preset threshold, the surgical approach is selected as the transabdominal approach.

7. The method for predicting the surgical difficulty of rectal cancer based on artificial intelligence according to claim 6, characterized in that, The acquisition of individual patient baseline characteristics includes: Determine the individual patient's physical numerical characteristics and treatment category characteristics; The normalized body numerical features and the embedded coded treatment category features are used as the baseline features of individual patients.

8. An artificial intelligence-based system for predicting the difficulty of rectal cancer surgery, characterized in that, include: The acquisition module is used to acquire high-resolution MRI images of the rectum. The segmentation module is used to call a semantic analysis model to segment the target region of pelvic anatomical structures from the high-resolution MRI image of the rectum; The extraction module is used to extract morphological features of pelvic anatomical structures based on three-dimensional point cloud data determined from the target region. The inference module is used to construct a graph structure model corresponding to the pelvic anatomical structure based on the three-dimensional point cloud data, and call the graph neural network to infer the graph structure model to obtain the graph features of the pelvic anatomical structure. The prediction module is used to call the surgical difficulty prediction model to predict the baseline features, graph features, and morphological features of an individual patient, so as to obtain the difficulty prediction results of rectal cancer surgery and the surgical approach auxiliary decision results.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the artificial intelligence-based method for predicting the difficulty of rectal cancer surgery as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the artificial intelligence-based method for predicting the difficulty of rectal cancer surgery as described in any one of claims 1 to 7.

Images