Segmentation method for key anatomical regions in pancreatic surgery based on ct images

Abstract

The present application relates to the technical field of image processing, in particular to a pancreas surgery key anatomical region segmentation method based on CT images, comprising the following steps: acquiring CT arterial phase images and CT venous phase images; preprocessing the CT arterial phase images and CT venous phase images in a window width window level adjustment and nonlinear filtering processing mode; through the combination of a TotalSegmentator segmentation model, an nnUNet segmentation framework and a region growing algorithm, key blood vessels such as a pancreaticoduodenal artery disconnection area and a PV-SMV axis are segmented from the preprocessed CT arterial phase images and CT venous phase images, and surrounding fat tissues are obtained. The present application realizes the automatic segmentation of key blood vessels such as the pancreaticoduodenal artery disconnection area and the PV-SMV axis by combining the nnNet segmentation framework and optimizing the TotalSegmentator technology, and obtains the surrounding fat gap, which not only solves the subjective deviation problem of traditional manual ROI delineation, but also provides a more stable and more surgery-demand-matching spatial framework for image-omics feature extraction by accurately positioning key anatomical areas such as the GDA area and the PV-SMV area.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method for segmenting key anatomical regions in pancreatic surgery based on CT images. Background Technology

[0002] Pancreaticoduodenectomy (PD) is the standard surgical procedure for treating lesions of the pancreatic head and periampullary region. The resection involves dense neurovascular plexuses (such as SMA, SMV / PV) within the mesopancreatic mesentery, making PD one of the most complex and challenging general surgical procedures. Especially when inflammation occurs in the pancreatic mesentery (such as in chronic pancreatitis or fibrosis following neoadjuvant therapy), the loose perivascular spaces are replaced by dense fibrous tissue, leading to the loss of normal perivascular anatomical layers and a significantly increased incidence of "difficult PD."

[0003] Preoperative contrast-enhanced CT is the primary diagnostic method for assessing inflammation in the surgical area, but the clinical guidance value of traditional CT findings is limited by technical limitations. Although CT images can clearly show the anatomical structure of the large vessels surrounding the pancreas, their low soft tissue resolution can lead to the failure to identify early microscopic inflammatory features (such as microcirculatory disturbances), thus underestimating the degree of inflammation in the pancreatic mesangial area during surgery. Furthermore, current imaging assessment strategies often focus on morphological changes in the "tumor-vessel contact area," neglecting the impact of the inflammatory response on the biomechanical properties of local tissues, which is precisely a key factor in predicting the difficulty of vascular dissection.

[0004] Radiomics extracts texture, morphology, and functional features from medical images through high-throughput processing, transforming microscopic tissue heterogeneity (such as the spatial distribution of inflammation-related fibrosis) into computable high-dimensional data. This has the potential to be used to construct clinically meaningful inflammation prediction models. Therefore, the construction of inflammation prediction models relies on the radiomic features of key areas in pancreatic surgery. Thus, the primary task is to accurately segment these key areas in CT images for radiomic feature extraction. Currently, manual delineation methods used for segmenting key areas in pancreatic surgery suffer from subjective bias. Automated delineation using AI algorithms cannot accurately locate critical anatomical regions such as the gastroduodenal artery (GDA) and the portal vein-superior mesenteric vein (PV-SMV), thus failing to provide a more stable and surgically appropriate spatial segmentation framework for radiomic feature extraction. Summary of the Invention

[0005] The purpose of this invention is to provide a method for segmenting key anatomical regions in pancreatic surgery based on CT images, in order to solve the technical problem that the existing technology cannot accurately locate key anatomical regions such as the gastroduodenal artery (GDA) and the portal vein-superior mesenteric vein (PV-SMV), thus failing to provide a more stable spatial segmentation framework that better meets surgical needs for the extraction of radiomics features.

[0006] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution:

[0007] A method for segmenting key anatomical regions in pancreatic surgery based on CT images, comprising the following steps:

[0008] Acquire CT arterial phase images and CT venous phase images;

[0009] The CT arterial phase images and CT venous phase images were preprocessed using window width and window level adjustment and nonlinear filtering.

[0010] By combining the TotalSegmentator segmentation model, the nnUNet segmentation framework, and the region growing algorithm, the following ROIs were segmented from the preprocessed CT venous phase images and the preprocessed CT arterial phase images: LPD surgical area connective tissue ROI 1, peripancreatic connective tissue ROI 2, connective tissue ROI 3 around the common hepatic artery and its main branches, connective tissue ROI 4 around the superior mesenteric artery, connective tissue ROI 5 around the SMV-PV axis, and connective tissue ROI 6 on the right side of the SMV-PV axis.

[0011] As a preferred embodiment of the present invention, the method for segmenting the connective tissue ROI 1 of the LPD surgical area includes:

[0012] On the pre-processed CT venous phase images, the important anatomical structures during LPD were segmented using the TotalSegmentator segmentation model, and the clipping box containing the pancreas, SMV-PV axis, celiac trunk, common hepatic artery and its main branches was determined.

[0013] On CT venous phase images within the clipping frame, abdominal visceral fat is locally segmented using the nnUNet segmentation framework;

[0014] On preprocessed CT venous phase images, abdominal visceral fat was globally segmented using the nnUnet segmentation framework.

[0015] Spatial registration and masking operations were performed on the local segmentation results of abdominal visceral fat and the global segmentation results of abdominal visceral fat to obtain the connective tissue ROI 1 of the LPD surgical area.

[0016] As a preferred embodiment of the present invention, the method for segmenting the peripancreatic connective tissue ROI 2 includes:

[0017] The overall contour of the pancreas was segmented using the TotalSegmentator segmentation model on the preprocessed CT venous phase images.

[0018] Based on the overall outline of the pancreas, extend 10 mm evenly into the surrounding three-dimensional space to generate the peripancreatic connective tissue ROI2.

[0019] The characterization of the peripancreatic connective tissue ROI 2 was optimized by setting the CT value range to -150HU to 60HU.

[0020] As a preferred embodiment of the present invention, the method for segmenting the connective tissue ROI 3 surrounding the common hepatic artery and its main branches and the connective tissue ROI 4 surrounding the superior mesenteric artery includes:

[0021] On the preprocessed CT arterial phase images, the lower boundary was manually determined, and the upper boundary was the lower edge of the 12th thoracic vertebra (i.e. the horizontal plane where the abdominal aorta originates). The left and right and anterior and posterior boundaries were determined according to the overall outline of the pancreas. A cubic region of interest containing the common hepatic artery and its main branches, as well as the superior mesenteric artery, was constructed. The main branches of the common hepatic artery include the proper hepatic artery and the initial segment of the gastroduodenal artery.

[0022] The pixel values ​​of the vertebral body, pancreas and aorta in the region of interest of the cube are set to zero. Taking advantage of the high density characteristics of blood vessels in CT images (CT value range: 150HU to 300HU), a region growing algorithm is used to trace the arterial tissue from the beginning of the abdominal aorta.

[0023] The region growing algorithm first identifies the first major branch originating from the anterior wall of the abdominal aorta: the celiac trunk, and then identifies the right-side branches of the celiac trunk: the common hepatic artery, the initial segment of the gastroduodenal artery, and the initial segment of the proper hepatic artery.

[0024] Then, the region growing algorithm identifies the second major branch of the abdominal aorta, the superior mesenteric artery, 1-2 cm below the celiac trunk.

[0025] The common hepatic artery, the initial segments of the gastroduodenal artery and the proper hepatic artery, and the superior mesenteric artery were expanded in three dimensions isotropically with an expansion radius of 10 mm to obtain the connective tissue ROI 3 around the initial segments of the common hepatic artery, the gastroduodenal artery and the proper hepatic artery, and the connective tissue ROI 4 around the superior mesenteric artery.

[0026] As a preferred embodiment of the present invention, the method for segmenting the connective tissue ROI 5 around the SMV-PV axis and the connective tissue ROI 6 on the right side of the SMV-PV axis includes:

[0027] On the preprocessed CT venous phase images, an initial three-dimensional region of interest containing the vein, portal vein, and splenic vein is constructed with the upper edge of the third lumbar vertebra (i.e., the plane where the horizontal part of the duodenum is located) as the lower boundary and the bifurcation of the portal vein as the upper boundary.

[0028] The pixel values ​​of the vertebral body, pancreas, superior mesenteric vein, portal vein, and splenic vein in the initial 3D region of interest are set to zero. Taking advantage of the high density characteristics of vascular tissue in CT venous phase images (CT value range: 150HU to 300HU), a region growing algorithm is used to track venous tissue starting from the bifurcation of the portal vein to identify the superior mesenteric vein, portal vein, and splenic vein.

[0029] The splenic vein region was removed 300 pixels to the left of the midline in the CT venous phase image to obtain the SMV-PV axis;

[0030] The SMV-PV axis was expanded in three dimensions isotropically with an expansion radius of 10 mm to obtain the ROI 5 of connective tissue around the SMV-PV axis;

[0031] Using the SMV-PV axis as the center, a region growth algorithm was used to identify the right half of the SMV-PV axis with a growth radius of 5 mm to obtain the connective tissue ROI 6 on the right side of the SMV-PV axis.

[0032] As a preferred embodiment of the present invention, the method for spatially registering and masking the local segmentation results of abdominal visceral fat with the global segmentation results of abdominal visceral fat includes:

[0033] The global segmentation result of abdominal visceral fat is spatially registered with the clipping box to obtain the global segmentation result of abdominal visceral fat in the clipping box coordinate system.

[0034] In the global segmentation results of abdominal visceral fat located in the clipping frame coordinate system, extract the global segmentation results of abdominal visceral fat within the clipping frame;

[0035] Real-time acquisition of the segmentation accuracy performance of the nnUNet segmentation framework for local segmentation of abdominal visceral fat, and real-time acquisition of the segmentation accuracy performance of the nnUNet segmentation framework for global segmentation of abdominal visceral fat.

[0036] Based on the segmentation accuracy of the local segmentation and the segmentation accuracy of the global segmentation, the global segmentation result of the abdominal visceral fat within the clipping frame and the local segmentation result of the abdominal visceral fat are subjected to a masking operation in a weighted fusion manner to obtain the connective tissue ROI 1 of the LPD surgical area.

[0037] The expression for the mask operation is:

[0038] ;

[0039] In the formula, The connective tissue of the LPD surgical area, This is the global segmentation result of abdominal visceral fat within the clipping frame. This is the result of local segmentation of abdominal visceral fat. For the accuracy of global segmentation, This is to assess the accuracy of local segmentation.

[0040] As a preferred embodiment of the present invention, the arterial and venous tissue tracking method of the region growing algorithm includes:

[0041] Step 1: Use the pancreatic mask frame as the seed for the entire vein, and use the upper and lower ends of the abdominal aorta and the L1-2 vertebral body as the seed for the entire artery;

[0042] Step 2: Apply morphological dilation operations to the overall vein and artery seeds to perform lateral dilation of the vein and artery tissues, while leaving the longitudinal ends undilated. The expression for the morphological dilation is:

[0043] ;

[0044] ;

[0045] In the formula, Ω represents the region after lateral expansion, x represents a point in the expanded region, y represents a point in the central axis portion, b represents the structural element point of the morphological expansion, and c(t) represents the central axis of the artery and vein. To exclude the areas near the ends of arteries and veins, only [the following is applied]. The constant term for the expansion of the centerline within the range, r is the expansion radius, C′ is the remaining centerline portion after removing the two endpoints of the arteries and veins, Br is the morphological structuring element of radius r, and ⊕ is the Minkowski operator;

[0046] Step 3: Determine the vascular spatial constraints and, based on these constraints, screen the growth regions of venous and arterial tissues. The expression for the vascular spatial constraints is:

[0047] ;

[0048] In the formula, Ω_vessel is the spatial region suitable for blood vessel growth, and D(x) is the characteristic attribute of point x in the expanded region. For a predefined direction vector, The threshold for the growth direction. Let x be the gray value of point x in the expanded region. and These are the minimum and maximum values ​​of the preset grayscale value range, respectively;

[0049] Step 4: Iteratively execute steps 2 and 3 to obtain the common hepatic artery, the initial segment of the gastroduodenal artery and the initial segment of the proper hepatic artery, the superior mesenteric artery, as well as the superior mesenteric vein, portal vein and splenic vein.

[0050] As a preferred embodiment of the present invention, the region growing algorithm identifies the right half of the SMV-PV axis with a growth radius of 5mm using a layered expansion method based on distance transformation and direction constraint. This layered expansion method based on distance transformation and direction constraint includes:

[0051] The binary mask of the SMV-PV axis is inverted, and the distance transformation algorithm is used to obtain the distance map from the pixels outside the SMV-PV axis to the nearest SMV-PV axis boundary;

[0052] On the distance map, all pixels not exceeding the SMV-PV axis boundary are first filtered out by a given 5mm threshold. The coordinates of the pixels are compared with the coordinates of their nearest blood vessel boundary point. Pixels within 5mm of the distance are only retained in the region to the right of the SMV-PV axis.

[0053] If the SMV-PV axis is curved and the definition of "right side" depends on the local normal, the center line of the SMV-PV axis is extracted first and the tangent vector and normal vector at each point are calculated. Then, among the candidate points that meet the 5mm threshold, the candidate points located within the "right normal" range are judged and retained to obtain the expansion region within 5mm to the right of the SMV-PV axis, which is taken as the connective tissue ROI 6 on the right side of the SMV-PV axis.

[0054] As a preferred embodiment of the present invention, the gold standard for segmentation of ROI 1 to ROI 6 is obtained manually by a radiologist with professional experience in abdominal imaging.

[0055] In a preferred embodiment of the present invention, the segmentation accuracy is quantified by the degree of difference between the global segmentation result and the local segmentation result of abdominal visceral fat within the clipping frame and the segmentation gold standard of ROI 1. The expression for quantifying the segmentation accuracy is as follows:

[0056] ;

[0057] ;

[0058] In the formula, The global segmentation result within the cropping box obtained by the nnUNet segmentation framework on the i-th image sample at the current time is given. The local segmentation result obtained by the nnUNet segmentation framework on the i-th image sample at the current time is... Let be the gold standard for segmentation of ROI 1 on the i-th image sample. Here is the mean squared loss formula, where m is the total number of image samples.

[0059] Compared with the prior art, the present invention has the following advantages:

[0060] This invention achieves automatic segmentation of key vessels such as the pancreatic and duodenal artery transection area and the PV-SMV axis by combining the nnNet segmentation framework with optimized TotalSegmentator technology, and obtains the surrounding fat space. This not only solves the subjective bias problem of traditional manual ROI delineation, but also provides a more stable and surgically-compliant spatial framework for radiomics feature extraction by accurately locating key anatomical areas such as the GDA and PV-SMV regions. Furthermore, this hierarchical modeling strategy, which combines the nnNet segmentation framework with TotalSegmentator technology to form a key region segmentation model, shifts the surgeon's visual evaluation standard from overall visual clarity to local anatomical resolution for the first time, providing a new technical path for accurate inflammation prediction. Attached Figure Description

[0061] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0062] Figure 1 A flowchart of a method for segmenting key anatomical regions in pancreatic surgery based on CT images, provided in an embodiment of the present invention;

[0063] Figure 2 A flowchart of the LPD surgical area connective tissue ROI 1 segmentation method provided in this embodiment of the invention;

[0064] Figure 3 This is a region-of-interest segmentation result diagram provided in an embodiment of the present invention;

[0065] Figure 4 This is a schematic diagram of the collected dataset provided in an embodiment of the present invention. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] like Figure 1 As shown, this invention provides a method for segmenting key anatomical regions in pancreatic surgery based on CT images, comprising the following steps:

[0068] Acquire CT arterial phase images and CT venous phase images;

[0069] CT arterial and venous phase images were preprocessed using window width and window level adjustment and nonlinear filtering. The window width was set to 350 Hounsfield Units (HU) to 400 HU, and the window level was set to 40 HU to 50 HU.

[0070] By combining the TotalSegmentator segmentation model, the nnUNet segmentation framework, and the region growing algorithm, the following ROIs were segmented from the preprocessed CT venous phase images and the preprocessed CT arterial phase images: LPD surgical area connective tissue ROI 1, peripancreatic connective tissue ROI 2, connective tissue ROI 3 around the common hepatic artery and its main branches, connective tissue ROI 4 around the superior mesenteric artery, connective tissue ROI 5 around the SMV-PV axis (superior mesenteric vein-portal vein axis), and connective tissue ROI 6 on the right side of the SMV-PV axis.

[0071] To achieve precise segmentation of key anatomical regions in pancreatic surgery, this invention constructs an artificial intelligence model that automatically segments the connective tissue ROI 1 of the LPD surgical area, the peripancreatic connective tissue ROI 2, the connective tissue ROI 3 surrounding the common hepatic artery and its main branches (proprioceptive hepatic artery and gastroduodenal artery), the connective tissue ROI 4 surrounding the superior mesenteric artery, the connective tissue ROI 5 surrounding the SMV-PV axis, and the connective tissue ROI 6 on the right side of the SMV-PV axis. This achieves automatic segmentation of key vessels such as the pancreaticoduodenal artery transection area and the PV-SMV axis, and obtains the surrounding fat spaces. This not only solves the subjective bias problem of traditional manual ROI delineation, but also provides a more stable and surgically-compliant spatial framework for radiomics feature extraction by precisely locating key anatomical regions such as the GDA region (gastroduodenal artery) and the PV-SMV region.

[0072] In constructing the artificial intelligence model, this invention employs a hierarchical modeling strategy that combines the TotalSegmentator segmentation model and the nnUNet segmentation framework to build a segmentation structure for key anatomical regions in pancreatic surgery. This allows the TotalSegmentator segmentation model to perform coarse segmentation of multiple organs, including the pancreas and major blood vessels, in advance. Then, the nnUNet segmentation framework performs fine segmentation of other key areas, such as the entire abdominal fat, based on the coarse segmentation, thereby improving segmentation accuracy while preserving information about multiple anatomical structures.

[0073] This invention incorporates a region growing algorithm into the pancreatic surgery key anatomical region segmentation structure, which is composed of the TotalSegmentator segmentation model and the nnUNet segmentation framework. This enables the comprehensive use of the TotalSegmentator segmentation model, the nnUNet segmentation framework, and the region growing algorithm to segment various key anatomical regions for pancreatic surgery, including the pancreas, the SMV-PV axis, the celiac trunk, the common hepatic artery and its main branches, and the surrounding fat spaces.

[0074] The TotalSegmentator segmentation model can identify blood vessel contours based on thresholds and simplify the blood vessel image into a skeleton line, where each pixel represents the center point of a blood vessel cross-section. After simplification, the endpoint or next bifurcation point of the blood vessel is found by identifying the topological structure of the skeleton line, thereby segmenting the blood vessel into multiple independent branches. Therefore, this invention can segment the pancreas, SMV-PV axis, celiac trunk, common hepatic artery and their main branch structures using the TotalSegmentator segmentation model, achieving accurate segmentation of key blood vessels such as the pancreatic-duodenal artery disconnection area and the PV-SMV axis.

[0075] The nnUNet segmentation framework is an adaptive deep learning framework specifically designed for medical image segmentation. Through an automated configuration process, it can achieve optimal performance in segmenting tumors or extracting organ contours on different datasets using the basic UNet architecture. Therefore, this invention applies it to the clipping boxes obtained by the TotalSegmentator segmentation model, which include the pancreas, SMV-PV axis, celiac trunk, common hepatic artery and its main branches. It can segment other key areas such as abdominal visceral fat, including LPD surgical area connective tissue ROI 1, peripancreatic connective tissue ROI 2, connective tissue ROI 3 around the common hepatic artery and its main branches, connective tissue ROI 4 around the superior mesenteric artery, connective tissue ROI 5 around the SMV-PV axis, and connective tissue ROI 6 on the right side of the SMV-PV axis.

[0076] Therefore, the hierarchical modeling strategy combining the TotalSegmentator segmentation model and the nnUNet segmentation framework can segment key vessels such as the pancreatic and duodenal artery transection area and the PV-SMV axis in CT images, while also segmenting the surrounding fat space, achieving a fusion of coarse and fine segmentation, preserving multiple anatomical structural information while improving segmentation accuracy.

[0077] Region growing algorithms perform image segmentation based on region features. Specifically, starting from a given or automatically selected seed point, pixels that meet the conditions within the region are repeatedly detected based on similarities in grayscale, curvature, and texture. After gradually merging adjacent pixels that meet the conditions, the target region is obtained.

[0078] This invention incorporates a region growing algorithm into the segmentation of abdominal visceral fat (connective tissue) and other key regions within the nnUNet segmentation framework. It uses a segmentation mask, cropped after locating blood vessels and the spine, as a whole "seed," without specifying individual seed points. All voxels currently within the mask are simultaneously considered as the starting region. This method of expansion and constraint selection based on the entire seed, rather than a region growing algorithm based on individual seed points, improves processing efficiency while maintaining the integrity of the expanded region and ensuring that the expanded area does not arbitrarily "intrude" into tissues that do not conform to HU values ​​or morphological characteristics.

[0079] In this invention, the nnUNet segmentation framework segments abdominal visceral tissues on a clipping frame containing the pancreas, SMV-PV axis, celiac trunk, common hepatic artery, and their main branches. This is a local image segmentation based on CT images, achieving refined segmentation focused on key regions. However, local image segmentation may inevitably lose some image information compared to global image segmentation, resulting in impaired segmentation accuracy. To compensate for the segmentation accuracy loss caused by this information loss, this invention performs abdominal visceral tissue segmentation on the original CT image using the nnUNet segmentation framework. This is a global image segmentation based on CT images. Through spatial registration and masking operations, the global image segmentation result is added to the local image segmentation result, which is equivalent to adding original image information to the local segmentation to reduce information loss. This avoids the segmentation accuracy loss caused by information loss and further improves the segmentation accuracy of the hierarchical modeling strategy.

[0080] This invention employs a hierarchical modeling strategy combined with the TotalSegmentator segmentation model (derived from TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CTImages. Radiology: Artificial Intelligence. https: / / doi.org / 10.1148 / ryai.230024.) and the nnUNet segmentation framework to construct segmentation structures for key anatomical regions in pancreatic surgery, such as... Figure 4 As shown, the dataset collection included: 112 patients without inflammation and 91 patients with inflammation, randomly assigned to the training and validation sets in a 7:3 ratio. 44 patients with mild inflammation and 47 patients with severe inflammation were also randomly assigned to the training and validation sets in a 7:3 ratio.

[0081] This invention comprehensively utilizes the TotalSegmentator segmentation model, the nnUNet segmentation framework, and region growing algorithms to segment key anatomical regions for pancreatic surgery, including the pancreas, SMV-PV axis, celiac trunk, common hepatic artery and its main branches, and surrounding fat spaces, as detailed below:

[0082] like Figure 2 As shown, the segmentation method for ROI 1 of connective tissue in the LPD surgical area includes:

[0083] On the pre-processed CT venous phase images, the important anatomical structures during LPD were segmented using the TotalSegmentator segmentation model, and the clipping box containing the pancreas, SMV-PV axis, celiac trunk, common hepatic artery and its main branches was determined.

[0084] On CT venous phase images within the clipping frame, abdominal visceral fat is locally segmented using the nnUNet segmentation framework;

[0085] On preprocessed CT venous phase images, abdominal visceral fat was globally segmented using the nnUnet segmentation framework.

[0086] Spatial registration and masking operations were performed on the local segmentation results of abdominal visceral fat and the global segmentation results of abdominal visceral fat to obtain the connective tissue ROI 1 of the LPD surgical area (denoted as Original when predicting the presence or absence of inflammation, and as V RegionFat when predicting mild and severe inflammation, e.g.) Figure 3 (as shown in a to c).

[0087] The segmentation methods for peripancreatic connective tissue ROI 2 include:

[0088] The overall contour of the pancreas was segmented using the TotalSegmentator segmentation model on the preprocessed CT venous phase images.

[0089] Based on the overall outline of the pancreas, extend uniformly 10 mm outwards into the surrounding three-dimensional space to generate the peripancreatic connective tissue ROI 2 (denoted as V Pancreas Original, e.g.) Figure 3 (as shown in d~f).

[0090] By setting the CT value range to -150 HU to 60 HU, the CT value range during peripancreatic connective tissue segmentation is also set to -150 HU to 60 HU. This CT value range covers normal connective tissue (CT value range: -150 HU to -30 HU) and inflammatory connective tissue (CT value range: -30 HU to 60 HU), thus more comprehensively reflecting the pathophysiological state of the peripancreatic connective tissue and further optimizing the characterization of the peripancreatic connective tissue ROI 2.

[0091] The segmentation methods for the connective tissue ROI 3 surrounding the common hepatic artery and its major branches, and the connective tissue ROI 4 surrounding the superior mesenteric artery include:

[0092] On the preprocessed CT arterial phase images, the lower boundary was manually determined, and the upper boundary was the lower edge of the 12th thoracic vertebra (i.e. the horizontal plane where the abdominal aorta originates). The left and right and anterior and posterior boundaries were determined according to the overall outline of the pancreas. A cubic region of interest containing the common hepatic artery and its main branches, as well as the superior mesenteric artery, was constructed. The main branches of the common hepatic artery include the proper hepatic artery and the initial segment of the gastroduodenal artery.

[0093] The pixel values ​​of the vertebral body, pancreas and aorta in the region of interest of the cube are set to zero. Taking advantage of the high density characteristics of blood vessels in CT images (CT value range: 150HU to 300HU), a region growing algorithm is used to trace the arterial tissue from the beginning of the abdominal aorta.

[0094] The region growing algorithm first identifies the first major branch originating from the anterior wall of the abdominal aorta: the celiac trunk, and then identifies the right-side branches of the celiac trunk: the common hepatic artery, the initial segment of the gastroduodenal artery, and the initial segment of the proper hepatic artery.

[0095] Then, the region growing algorithm identifies the second major branch of the abdominal aorta, the superior mesenteric artery, 1-2 cm below the celiac trunk.

[0096] The common hepatic artery, the initial segments of the gastroduodenal artery and the proper hepatic artery, and the superior mesenteric artery were expanded three-dimensionally isotropically with an extension radius of 10 mm to obtain the ROI 3 (denoted as A1 Original) of connective tissue surrounding the initial segments of the common hepatic artery, the gastroduodenal artery, and the proper hepatic artery. Figure 3 (as shown in g) and the connective tissue ROI 4 around the superior mesenteric artery (denoted as A2 Original, as shown in g) Figure 3 (As shown in g). This invention achieves accurate identification of the connective tissue surrounding the main blood supply arteries of the pancreatic head in the pancreatic mesangial region through the above method.

[0097] The segmentation methods for ROI 5 of the connective tissue around the SMV-PV axis and ROI 6 of the connective tissue to the right of the SMV-PV axis include:

[0098] On the preprocessed CT venous phase images, an initial three-dimensional region of interest containing the vein, portal vein, and splenic vein is constructed with the upper edge of the third lumbar vertebra (i.e., the plane where the horizontal part of the duodenum is located) as the lower boundary and the bifurcation of the portal vein as the upper boundary.

[0099] The pixel values ​​of the vertebral body, pancreas, superior mesenteric vein, portal vein, and splenic vein in the initial 3D region of interest are set to zero. Taking advantage of the high density characteristics of vascular tissue in CT venous phase images (CT value range: 150HU to 300HU), a region growing algorithm is used to track venous tissue starting from the bifurcation of the portal vein to identify the superior mesenteric vein, portal vein, and splenic vein.

[0100] The splenic vein region was cropped 300 pixels to the left of the central axis in the CT venous phase image to obtain the SMV-PV axis. Since the splenic vein was not the target vessel in this study, it was cropped 300 pixels to the left of the central axis in the image (the original image resolution was 512×512). The SMV-PV axis was obtained after cropping the splenic vein region.

[0101] The SMV-PV axis was isotropically extended in three dimensions with an extension radius of 10 mm to obtain the connective tissue ROI 5 (denoted as V1 Original) around the SMV-PV axis. Figure 3 (as shown in g)

[0102] Centered on the SMV-PV axis, a region growth algorithm was used to identify the right half of the SMV-PV axis with a growth radius of 5 mm to obtain the connective tissue ROI 6 on the right side of the SMV-PV axis (denoted as V1rightoriginal).

[0103] In this invention, the nnUNet segmentation framework segments abdominal visceral tissues on a clipping box containing the pancreas, SMV-PV axis, celiac trunk, common hepatic artery, and their main branches. This is a local image segmentation based on CT images, achieving refined segmentation focused on key regions. However, local image segmentation, compared to global image segmentation, may inevitably lose some image information, resulting in compromised segmentation accuracy. To compensate for this information loss, this invention performs a second segmentation of abdominal visceral tissues on the original CT image using the nnUNet framework. This is a global image segmentation based on the CT image. Through spatial registration and masking operations, the global image segmentation result is added to the local image segmentation result, effectively adding original image information to the local segmentation to reduce information loss and avoid the segmentation accuracy loss caused by information loss. This further improves the segmentation accuracy of the hierarchical modeling strategy, as detailed below:

[0104] Methods for spatial registration and masking operations between local and global segmentation results of abdominal visceral fat include:

[0105] The global segmentation result of abdominal visceral fat is spatially registered with the clipping box to obtain the global segmentation result of abdominal visceral fat in the clipping box coordinate system.

[0106] In the global segmentation results of abdominal visceral fat located in the clipping frame coordinate system, extract the global segmentation results of abdominal visceral fat within the clipping frame;

[0107] Real-time acquisition of the segmentation accuracy performance of the nnUNet segmentation framework for local segmentation of abdominal visceral fat, and real-time acquisition of the segmentation accuracy performance of the nnUNet segmentation framework for global segmentation of abdominal visceral fat.

[0108] Based on the segmentation accuracy of local segmentation and global segmentation, the global segmentation results of abdominal visceral fat within the clipping frame and the local segmentation results of abdominal visceral fat were subjected to a weighted fusion masking operation to obtain the connective tissue ROI 1 of the LPD surgical area.

[0109] The expression for mask operation is:

[0110] ;

[0111] In the formula, Connective tissue in the LPD surgical area, This is the global segmentation result of abdominal visceral fat within the clipping frame. This is the result of local segmentation of abdominal visceral fat. For the accuracy of global segmentation, This is to assess the accuracy of local segmentation.

[0112] This invention employs a weighted fusion approach in masking operations, introducing the nnUNet segmentation framework to assess the real-time segmentation accuracy of both local and global segmentation. This involves evaluating the performance of the global and local segmentation models within the nnUNet framework at the current moment; higher performance equates to higher confidence in the segmentation result. Therefore, the normalized results serve as weights for both the global and local segmentation results. This allows the lost segmentation accuracy from local segmentation to be supplemented by global segmentation, achieving the goal of "adding global image segmentation results to local image segmentation results, which is equivalent to adding original image information to local segmentation to reduce information loss, thus avoiding the segmentation accuracy loss caused by information loss and further improving the segmentation accuracy of the hierarchical modeling strategy."

[0113] The method for tracing arterial and venous tissues using the region growing algorithm includes: Step 1: Using pancreatic mask frame clipping as the seed for the entire vein, and using the abdominal aorta and the upper and lower ends of the L1~2 vertebral body clipping of the abdominal aorta as the seed for the entire artery;

[0114] Step 2: Apply morphological dilation operations to the overall vein and artery seeds to perform lateral dilation of the vein and artery tissues, while leaving the longitudinal ends undilated. The expression for morphological dilation is:

[0115] ;

[0116] ;

[0117] In the formula, Ω represents the region after lateral expansion, x represents a point in the expanded region, y represents a point in the central axis portion, b represents the structural element point of the morphological expansion, and c(t) represents the central axis of the artery and vein. To exclude the areas near the ends of arteries and veins, only [the following is applied]. The constant term for expanding the centerline within the range, r is the expansion radius or the size of the structuring element, that is, the distance of the lateral "expansion" of the centerline, C′ is the remaining part of the centerline after removing the two ends of the arteries and veins, Br is the morphological structuring element with radius r, and ⊕ is the Minkowski operator.

[0118] Centerline parameterization: Suppose that the centerline of a blood vessel can be represented by one (or more) curves. For a continuous curve c(t), where... This represents the parameters from the starting point to the ending point. The starting point is c(0), and the ending point is c(1). To "remove endpoints," only consider... And ignore the endpoints corresponding to t=0 and t=1 ( It can be set to a very small positive number as needed, or it can be set to 0 for special handling during expansion.

[0119] Define the expansion radius (or structuring element): Typically, the blood vessel is expanded in the "lateral" direction, i.e., the region at a distance r on the curve. For example, in a two-dimensional case, a disk Br (a circle with radius r) is used as the structuring element.

[0120] Step 3: Determine the spatial constraints of blood vessels, and screen the growth regions of venous and arterial tissues based on these constraints. The expression for the spatial constraints of blood vessels is:

[0121] ;

[0122] In the formula, Ω_vessel is the spatial region suitable for blood vessel growth, and D(x) is the characteristic attribute of point x in the expanded region. For a predefined direction vector, The threshold for the growth direction. Let x be the gray value of point x in the expanded region. and These are the minimum and maximum values ​​of the preset grayscale value range, respectively;

[0123] The spatial constraints of blood vessels in this invention are based on prior anatomical knowledge. Based on the clipping of relative positions, the spatial constraints of blood vessel growth are also based on prior anatomical knowledge, including two priors: the first is the direction of blood vessels, and the second is the HU value intensity of blood vessels.

[0124] Ω_vessel represents a suitable spatial region for blood vessel growth; x belongs to the growth point to be evaluated. The first constraint, D(x)·v_prior≥θ_dir, ensures that the relationship between the feature attribute D(x) of point x and the predefined direction vector v_prior conforms to the growth direction threshold θ_dir, thus filtering out suitable growth directions. The second constraint, HU_min≤I(x)≤HU_max, limits the gray value I(x) of point x to a reasonable range, ensuring that only tissues that conform to vascular characteristics are selected. Through these constraints, growth regions can be effectively screened for venous and arterial tissues.

[0125] Step 4: Iteratively execute steps 2 and 3 to obtain the common hepatic artery, the initial segment of the gastroduodenal artery and the initial segment of the proper hepatic artery, the superior mesenteric artery, as well as the superior mesenteric vein, portal vein and splenic vein.

[0126] This invention uses an initial mask, which is located and cut through blood vessels and spine, as the overall "seed" without specifying a single seed point, and considers all voxels currently in the mask as the starting region at the same time.

[0127] Morphological dilation: Perform dilation operation on the mask, expanding outward based on the pixel / voxel neighborhood; the mask can be preprocessed as needed before dilation (such as noise reduction, thinning, closing small holes, etc.) to improve the accuracy of subsequent growth.

[0128] Constraint screening: During the expansion process, check whether each candidate voxel in the expansion region meets the specified constraints.

[0129] HU value range: If the CT value of a voxel exceeds the reasonable range for blood vessels or target tissues, it will not be incorporated into the new mask;

[0130] Orientation / neighborhood features: By combining vascular geometry, gradient information, or other morphological features, regions that expand in unreasonable directions can be filtered out.

[0131] Only voxels that meet all the constraints will be incorporated into the final "expanded" vascular segmentation mask.

[0132] Iterative or multi-step expansion: Usually, expansion is not performed only once, but rather first expanded and screened under a smaller expansion kernel, and then iterated several times to gradually obtain a more accurate boundary; if the blood vessel edge is relatively smooth, or the HU value changes significantly, a stable state can often be reached with only a few iterations.

[0133] By using this method of expansion and constraint screening based on the whole seed block, rather than the region growth algorithm based on individual seed points, we can improve processing efficiency, maintain the integrity of the expanded region, and ensure that the expansion range does not arbitrarily "intrude" into tissues that do not conform to HU values ​​or morphological characteristics.

[0134] The region growing algorithm identifies the right half of the SMV-PV axis with a growth radius of 5mm. It employs a layered expansion method based on distance transformation and orientation constraint. The layered expansion method based on distance transformation and orientation constraint includes:

[0135] The binary mask of the SMV-PV axis is inverted, and the distance transformation algorithm is used to obtain the distance map from the pixels outside the SMV-PV axis to the nearest SMV-PV axis boundary;

[0136] On the distance map, all pixels not exceeding the SMV-PV axis boundary are first filtered out by a given 5mm threshold. The coordinates of the pixel are compared with the coordinates of its nearest blood vessel boundary point. Pixels within the 5mm threshold are only retained in the region with "larger x coordinate" (to the right of the SMV-PV axis).

[0137] If the SMV-PV axis is curved and the definition of "right side" depends on the local normal, the center line of the SMV-PV axis is extracted first and the tangent vector and normal vector at each point are calculated. Then, among the candidate points that meet the 5mm threshold, the candidate points located within the "right normal" range are judged and retained to obtain the expansion region within 5mm to the right of the SMV-PV axis, which is taken as the connective tissue ROI 6 on the right side of the SMV-PV axis.

[0138] In order to retain only the expansion effect on the "right side" of the blood vessel, this invention needs to apply directional restrictions by combining coordinate information or local normals: If only the "positive x-axis direction" in the image coordinate system is considered, the coordinates of the pixel can be compared with the coordinates of its nearest blood vessel boundary point, and only those pixels with "larger x-coordinates" are retained for pixels within 5mm of the distance; if the blood vessel is curved and the definition of "right side" depends on local normals, the center line of the blood vessel must first be extracted and the tangent vector and normal vector at each point must be calculated. Then, among the candidate points whose distance meets the threshold, it is further determined whether they are within the range of the "right normal" at that point, so as to obtain the final expansion area within 5mm to the right of the blood vessel.

[0139] In this context, a local normal is a vector perpendicular to the tangent vector at a specific point on a curve or surface. When dealing with the SMV-PV axis, the local normal is used to determine the left-right direction at the curve's bend. Mathematically, given a point P on a curve, its tangent vector T represents the direction of the curve at point P, while the normal vector N is a vector perpendicular to the tangent direction, indicating whether it's closer to the inside or outside of the curve.

[0140] The right-side normal is a special case of the local normal, typically defined by the direction of the tangent vector. On the SMV-PV axis, the right-side normal refers to the region formed to the right of the tangent vector by the normal vector. By calculating the normal vector at each point, it is possible to identify which candidate points lie in the relative region to the right of the SMV-PV axis. This is crucial for accurately identifying and expanding the perivascular connective tissue. The gold standard for segmentation of ROIs 1 through ROI 6 was manually segmented by radiologists with expertise in abdominal imaging.

[0141] In this invention, the real-time segmentation accuracy of the nnUNet segmentation framework for local and global segmentation is measured using a training set obtained from collected case samples. Specifically, the error between the local and global segmentation results obtained by the nnUNet segmentation framework in the training set and the gold standard segmentation of ROI 1 obtained manually by a radiologist with expertise in abdominal imaging is calculated. The higher the error, the lower the segmentation performance and the lower the confidence level of the segmentation results.

[0142] Segmentation accuracy is quantified by the degree of difference between the global segmentation results and the local segmentation results of abdominal visceral fat within the clipping frame and the gold standard segmentation of ROI 1. The expression for quantifying segmentation accuracy is:

[0143] ;

[0144] ;

[0145] In the formula, This represents the global segmentation result within the cropping box obtained by the nnUNet segmentation framework on the i-th image sample at the current time. This represents the local segmentation result obtained by the nnUNet segmentation framework on the i-th image sample at the current time. Let be the gold standard for segmentation of ROI 1 on the i-th image sample. This is the mean squared loss formula, where m is the total number of image samples, consistent with the number of cases in the training set.

[0146] In summary, combining the TotalSegmentator segmentation model with nnUNet enables a more comprehensive, accurate, and automated workflow for segmenting key pancreatic regions, primarily in the following three aspects:

[0147] 1. Synergy between coarse and fine segmentation of multiple structures: The TotalSegmentator segmentation model performs coarse segmentation of multiple organs, including target regions such as the pancreas and large blood vessels, in advance; nnUNet then performs fine segmentation of other key regions such as the entire abdominal fat, thereby improving segmentation accuracy while preserving information on multiple anatomical structures.

[0148] 2. Automated configuration and high scalability: nnUNet has built-in automated data preprocessing, network structure selection and hyperparameter tuning, which can significantly reduce model development and debugging time; the TotalSegmentator segmentation model provides a stable multi-organ segmentation framework and pre-trained weights, without the need for tedious manual settings, which is convenient for introducing new anatomical targets or extending and tuning existing target structures.

[0149] 3. Enhanced robustness and generalization ability: nnUNet effectively addresses data differences across multiple centers and scanning protocols through mechanisms such as automatic normalization, data augmentation, and post-processing; the TotalSegmentator segmentation model, based on large-scale, multi-structured training data, exhibits strong adaptability to diverse clinical images and reduces the impact of domain transformation. This provides more accurate surgical planning and tissue identification data for clinical practice and research.

[0150] This invention achieves automatic segmentation of key vessels such as the pancreatic-duodenal artery transection area and the PV-SMV axis by combining the nnNet segmentation framework with optimized TotalSegmentator technology, and obtains the surrounding fat space. This not only solves the subjective bias problem of traditional manual ROI delineation, but also provides a more stable and surgically-compliant spatial framework for radiomics feature extraction by accurately locating key anatomical areas such as the GDA and PV-SMV regions. Furthermore, this hierarchical modeling strategy, which combines the nnNet segmentation framework with TotalSegmentator technology to form a key region segmentation model, shifts the surgeon's visual evaluation standard from overall visual clarity to local anatomical resolution for the first time, providing a new technical path for accurate inflammation prediction.

[0151] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A method for segmenting key anatomical regions in pancreatic surgery based on CT images, characterized in that, Includes the following steps: Acquire CT arterial phase images and CT venous phase images; The CT arterial phase images and CT venous phase images were preprocessed using window width and window level adjustment and nonlinear filtering. By combining the TotalSegmentator segmentation model, the nnUNet segmentation framework, and the region growing algorithm, the following ROIs were segmented from the preprocessed CT venous phase images and the preprocessed CT arterial phase images: LPD surgical area connective tissue ROI 1, peripancreatic connective tissue ROI 2, connective tissue ROI 3 around the common hepatic artery and its main branches, connective tissue ROI 4 around the superior mesenteric artery, connective tissue ROI 5 around the SMV-PV axis, and connective tissue ROI 6 on the right side of the SMV-PV axis. The segmentation methods for the connective tissue ROI 3 surrounding the common hepatic artery and its main branches and the connective tissue ROI 4 surrounding the superior mesenteric artery include: On the preprocessed CT arterial phase images, the lower boundary was manually determined, the lower edge of the 12th thoracic vertebra was used as the upper boundary, and the left, right and anterior and posterior boundaries were determined according to the overall outline of the pancreas. A cubic region of interest containing the common hepatic artery and its main branches, and the superior mesenteric artery was constructed. The main branches of the common hepatic artery include the proper hepatic artery and the initial segment of the gastroduodenal artery. The pixel values ​​of the vertebral body, pancreas and aorta in the region of interest of the cube are set to zero. Taking advantage of the high density of blood vessels in CT images, a region growing algorithm is used to trace the arterial tissue from the beginning of the abdominal aorta. The region growing algorithm first identifies the first major branch originating from the anterior wall of the abdominal aorta: the celiac trunk, and then identifies the right-side branches of the celiac trunk: the common hepatic artery, the initial segment of the gastroduodenal artery, and the initial segment of the proper hepatic artery. Then, the region growing algorithm identifies the second major branch of the abdominal aorta, the superior mesenteric artery, 1-2 cm below the celiac trunk. The common hepatic artery, the initial segments of the gastroduodenal artery and the proper hepatic artery, and the superior mesenteric artery were expanded in three dimensions isotropically with an expansion radius of 10 mm to obtain the connective tissue ROI 3 around the initial segments of the common hepatic artery, the gastroduodenal artery and the proper hepatic artery, and the connective tissue ROI 4 around the superior mesenteric artery.

2. The method for segmenting key anatomical regions in pancreatic surgery based on CT images according to claim 1, characterized in that: The segmentation method for the connective tissue ROI 1 of the LPD surgical area includes: On the pre-processed CT venous phase images, the important anatomical structures during LPD were segmented using the TotalSegmentator segmentation model, and the clipping box containing the pancreas, SMV-PV axis, celiac trunk, common hepatic artery and its main branches was determined. On CT venous phase images within the clipping frame, abdominal visceral fat is locally segmented using the nnUNet segmentation framework; On preprocessed CT venous phase images, abdominal visceral fat was globally segmented using the nnUnet segmentation framework. Spatial registration and masking operations were performed on the local segmentation results of abdominal visceral fat and the global segmentation results of abdominal visceral fat to obtain the connective tissue ROI 1 of the LPD surgical area.

3. The method for segmenting key anatomical regions in pancreatic surgery based on CT images according to claim 1, characterized in that: The segmentation method for the peripancreatic connective tissue ROI 2 includes: The overall contour of the pancreas was segmented using the TotalSegmentator segmentation model on the preprocessed CT venous phase images. Based on the overall outline of the pancreas, extend 10 mm evenly into the surrounding three-dimensional space to generate the peripancreatic connective tissue ROI 2; The characterization of the peripancreatic connective tissue ROI 2 was optimized by setting the CT value range to -150HU to 60HU.

4. The method for segmenting key anatomical regions in pancreatic surgery based on CT images according to claim 1, characterized in that: The segmentation method for the connective tissue ROI 5 around the SMV-PV axis and the connective tissue ROI 6 to the right of the SMV-PV axis includes: On the preprocessed CT venous phase images, an initial three-dimensional region of interest containing the vein, portal vein, and splenic vein is constructed with the upper edge of the third lumbar vertebra as the lower boundary and the bifurcation of the portal vein as the upper boundary. The pixel values ​​of the vertebral body, pancreas, superior mesenteric vein, portal vein, and splenic vein in the initial three-dimensional region of interest are set to zero. Taking advantage of the high density of vascular tissue in CT venous phase images, a region growing algorithm is used to track venous tissue starting from the bifurcation of the portal vein to identify the superior mesenteric vein, portal vein, and splenic vein. The splenic vein region was removed 300 pixels to the left of the midline in the CT venous phase image to obtain the SMV-PV axis; The SMV-PV axis was expanded in three dimensions isotropically with an expansion radius of 10 mm to obtain the ROI5 of connective tissue around the SMV-PV axis. Using the SMV-PV axis as the center, a region growth algorithm was used to identify the right half of the SMV-PV axis with a growth radius of 5 mm to obtain the connective tissue ROI 6 on the right side of the SMV-PV axis.

5. The method for segmenting key anatomical regions in pancreatic surgery based on CT images according to claim 2, characterized in that: The method for spatially registering and masking the local segmentation results of abdominal visceral fat with the global segmentation results of abdominal visceral fat includes: The global segmentation result of abdominal visceral fat is spatially registered with the clipping box to obtain the global segmentation result of abdominal visceral fat in the clipping box coordinate system. In the global segmentation results of abdominal visceral fat located in the clipping frame coordinate system, extract the global segmentation results of abdominal visceral fat within the clipping frame; Real-time acquisition of the segmentation accuracy performance of the nnUNet segmentation framework for local segmentation of abdominal visceral fat, and real-time acquisition of the segmentation accuracy performance of the nnUNet segmentation framework for global segmentation of abdominal visceral fat. Based on the segmentation accuracy of the local segmentation and the segmentation accuracy of the global segmentation, the global segmentation result of the abdominal visceral fat within the clipping frame and the local segmentation result of the abdominal visceral fat are subjected to a masking operation in a weighted fusion manner to obtain the connective tissue ROI 1 of the LPD surgical area. The expression for the mask operation is: In the formula, The connective tissue of the LPD surgical area, This is the global segmentation result of abdominal visceral fat within the clipping frame. This is the result of local segmentation of abdominal visceral fat. For the accuracy of global segmentation, This is to assess the accuracy of local segmentation.

6. The method for segmenting key anatomical regions in pancreatic surgery based on CT images according to claim 1, characterized in that: The arterial and venous tissue tracking methods of the region growing algorithm include: Step 1: Use the pancreatic mask frame as the seed for the entire vein, and use the upper and lower ends of the abdominal aorta and the L1-2 vertebral body as the seed for the entire artery; Step 2: Apply morphological dilation operations to the overall vein and artery seeds to perform lateral dilation of the vein and artery tissues, while leaving the longitudinal ends undilated. The expression for the morphological dilation is: In the formula, x is a point in the expanded region, y is a point in the central line portion, b is a morphological expansion structuring element point, c(t) is the central line of the artery and vein, r is the expansion radius, C′ is the remaining central line portion after removing the two endpoints of the artery and vein, Br is a morphological structuring element with radius r, and ⊕ is the Minkowski operator. Step 3: Determine the vascular spatial constraints and, based on these constraints, screen the growth regions of venous and arterial tissues. The expression for the vascular spatial constraints is: ； In the formula, Ω_vessel is the spatial region suitable for blood vessel growth, and D(x) is the characteristic attribute of point x in the expanded region. For a predefined direction vector, The threshold for the growth direction. Let x be the gray value of point x in the expanded region. and These are the minimum and maximum values ​​of the preset grayscale value range, respectively; Step 4: Iteratively execute steps 2 and 3 to obtain the common hepatic artery, the initial segment of the gastroduodenal artery and the initial segment of the proper hepatic artery, the superior mesenteric artery, as well as the superior mesenteric vein, portal vein and splenic vein.

7. The method for segmenting key anatomical regions in pancreatic surgery based on CT images according to claim 4, characterized in that: The region growing algorithm identifies the right half of the SMV-PV axis with a growth radius of 5mm using a layered expansion method based on distance transformation and direction constraint. This layered expansion method based on distance transformation and direction constraint includes: The binary mask of the SMV-PV axis is inverted, and the distance transformation algorithm is used to obtain the distance map from the pixels outside the SMV-PV axis to the nearest SMV-PV axis boundary; On the distance map, all pixels not exceeding the SMV-PV axis boundary are first filtered out by a given 5mm threshold. The coordinates of the pixel are compared with the coordinates of its nearest blood vessel boundary point. For pixels within 5mm of the distance, only the region located to the right of the SMV-PV axis is retained. If the SMV-PV axis is curved and the definition of "right side" depends on the local normal, the center line of the SMV-PV axis is extracted first and the tangent vector and normal vector at each point are calculated. Then, among the candidate points that meet the 5mm threshold, the candidate points located within the right normal range at that point are judged and retained to obtain the expansion region within 5mm to the right of the SMV-PV axis, which is taken as the connective tissue ROI 6 on the right side of the SMV-PV axis.

8. The method for segmenting key anatomical regions in pancreatic surgery based on CT images according to claim 1, characterized in that: The gold standard for segmentation of ROI 1 to ROI 6 was obtained manually by radiologists with expertise in abdominal imaging.

9. A method for segmenting key anatomical regions in pancreatic surgery based on CT images according to claim 5, characterized in that: The segmentation accuracy is quantified by the degree of difference between the global segmentation result and the local segmentation result of abdominal visceral fat within the clipping frame and the segmentation gold standard of ROI 1. The expression for quantifying the segmentation accuracy is: ； ； In the formula, The global segmentation result within the cropping box obtained by the nnUNet segmentation framework on the i-th image sample at the current time is given. The local segmentation result obtained by the nnUNet segmentation framework on the i-th image sample at the current time is... Let be the gold standard for segmentation of ROI 1 on the i-th image sample. Here is the mean squared loss formula, where m is the total number of image samples.

Images