Method for discriminating presence of pancreatic surgical field inflammation fused with fat-related image features

Abstract

The present application relates to the technical field of medical imageomics analysis, in particular to a pancreatic surgical area inflammation existence discrimination method fusing fat-related image features, comprising the following steps: determining clinical risk factors by statistical methods; on CT venous phase images, segmenting LPD surgical area connective tissue ROI 1 by combining a TotalSegmentator segmentation model and an nnUNet segmentation framework, and extracting ROI 1 imageomics features on the LPD surgical area connective tissue ROI 1; and constructing an inflammation existence discrimination model based on the ROI 1 imageomics features and clinical risk factors by using multiple machine learning algorithms. In the fusion model constructed by combining risk factors and imageomics features, the present application uses a retrospective study to find clinical risk factors, has screening operation support, and at the same time, the clinical risk factors and imageomics features are spliced and processed by using an attention mechanism, so that the fusion model quickly captures key features, thereby ensuring the inflammation existence discrimination accuracy and efficiency of the fusion model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical imageomics analysis, in particular to a pancreatic surgical area inflammation existence discrimination method fusing fat-related image features. BACKGROUND

[0002] Pancreaticoduodenectomy (PD) is the standard surgical procedure for treating lesions around the head of the pancreas and the ampulla, and its resection range involves the dense vascular nerve plexus (such as SMA, SMV / PV, etc.) in the mesopancreas. Therefore, PD is considered to be one of the most complex and challenging general surgery operations. Especially when the mesopancreas area has an inflammatory reaction (such as chronic pancreatitis, fibrosis after neoadjuvant therapy), the loose space around the blood vessels is replaced by dense fibrous tissue, resulting in the disappearance of the normal anatomic level around the blood vessels, and the incidence of "difficult PD" significantly increases.

[0003] Preoperative enhanced CT is the main examination method for evaluating the inflammation of the surgical area, but the traditional CT signs are limited by technical limitations, which restricts their clinical guidance value. Although CT images can clearly show the anatomic structure of the large blood vessels around the pancreas, the low soft tissue resolution of CT images cannot identify the early microscopic inflammatory features (such as microcirculation disorders), which may underestimate the degree of inflammation in the mesopancreas area during surgery. In addition, existing imaging evaluation strategies focus on the morphological changes in the "tumor-vascular contact range", ignoring the influence of the inflammatory response on the mechanical properties of the local tissue, which is actually a key factor in predicting the difficulty of blood vessel separation.

[0004] Imageomics extracts texture, morphology and functional features from medical images through high-throughput, and converts microscopic tissue heterogeneity (such as inflammation-related fibrosis spatial distribution) into computable high-dimensional data, which has the potential to be used to build an inflammation prediction model with clinical significance. Therefore, the current judgment of whether the surgical area is inflamed by imageomics has high clinical value, and even a fusion model is constructed by combining clinical risk factors to fuse multi-modal features for inflammation existence discrimination. However, in the process of constructing the fusion model, the selection of clinical risk factors and imageomics features is usually selected by expert subjective experience, lacking objective screening operations, thereby affecting the inflammation existence discrimination accuracy of the fusion model, and the clinical risk factors and imageomics features are usually directly spliced, which is difficult to make the fusion model quickly capture key features, thereby affecting the inflammation existence discrimination efficiency of the fusion model. SUMMARY

[0005] The application aims to provide a pancreatic surgery area inflammation existence discrimination method fusing fat-related image features, so as to solve the technical problems that in the prior art, the screening of clinical risk factors and image features of a clinical-imageomics fusion model is usually selected by subjective experience of experts, which affects the inflammation existence discrimination accuracy of the fusion model, and the clinical risk factors and image features are usually directly spliced, which is difficult to make the fusion model quickly capture key features, and affects the inflammation existence discrimination efficiency of the fusion model.

[0006] To solve the above technical problems, the application specifically provides the following technical solutions:

[0007] A pancreatic surgery area inflammation existence discrimination method fusing fat-related image features, comprising the following steps:

[0008] Determine clinical risk factors for pancreatic surgery area inflammation existence discrimination by statistical methods;

[0009] On a CT venous phase image, segment LPD surgery area connective tissue ROI 1 by a TotalSegmentator segmentation model and an nnUNet segmentation framework, and extract ROI 1 image features for pancreatic surgery area inflammation existence discrimination on the LPD surgery area connective tissue ROI 1;

[0010] Construct an inflammation existence discrimination model for pancreatic surgery area inflammation existence discrimination by various machine learning algorithms based on the ROI 1 image features and the clinical risk factors.

[0011] As a preferred solution of the application, the determination method of the clinical risk factors comprises:

[0012] Obtain basic clinical data, laboratory examination data, preoperative abdominal enhanced CT data, pathological data and perioperative outcome data;

[0013] Determine whether the measurement data in the basic clinical data, laboratory examination data, preoperative abdominal enhanced CT data, pathological data and perioperative outcome data conform to normal distribution by Shapiro-Wilk test, wherein:

[0014] Use mean and standard deviation to describe the continuous variables in the measurement data conforming to normal distribution, and use two independent sample t test when the variances of the two groups are equal, otherwise use Mann-Whitney U test;

[0015] Use median, 25th percentile and 75th percentile to describe the variables in the measurement data not conforming to non-normal distribution, and use Mann-Whitney U test for comparison between groups;

[0016] The basic clinical data, laboratory examination data, preoperative abdominal enhanced CT data, pathological data, and the counting data in the perioperative outcome data are described as number and percentage, and the chi-square test is used for comparison between groups, and the Fisher exact test is used for comparison between groups when the theoretical frequency exists, and the continuity correction chi-square test is used when 1 < theoretical frequency < 5 exists;

[0017] The Fisher exact test is used for comparison between groups when the theoretical frequency exists, and the continuity correction chi-square test is used when 1 < theoretical frequency < 5 exists;

[0018] When there is no correction formula for comparison among multiple groups, the Fisher exact test is used when the theoretical frequency < 1 or the number of cells with 1 ≤ theoretical frequency < 5 exceeds 20% of the total number of cells;

[0019] Stepwise regression method is used to screen variables with statistical difference in single factor analysis, and binary logistic regression analysis is used to determine clinical risk factors.

[0020] As a preferred scheme of the present application, the segmentation method of the LPD operation area connective tissue ROI 1 comprises:

[0021] The CT venous phase image is preprocessed in a window width and window level adjustment and nonlinear filtering processing mode;

[0022] On the preprocessed CT venous phase image, the LPD intraoperative important anatomical structure is segmented out by the TotalSegmentator segmentation model to determine the clipping frame containing the pancreas, SMV-PV axis, celiac trunk, common hepatic artery and its main branch structure;

[0023] On the CT venous phase image in the clipping frame, the abdominal visceral fat is locally segmented by the nnUNet segmentation framework;

[0024] On the preprocessed CT venous phase image, the abdominal visceral fat is globally segmented by the nnUnet segmentation framework;

[0025] The local segmentation result of the abdominal visceral fat and the global segmentation result of the abdominal visceral fat are spatially registered and mask operated to obtain the LPD operation area connective tissue ROI 1.

[0026] As a preferred scheme of the present application, the extraction method of the ROI 1 imageomics features comprises:

[0027] 107 imageomics features are extracted in the LPD operation area connective tissue ROI 1 by the PyRadiomics component in the Python library;

[0028] The 20 important imaging features consisting of 1 shape feature, 4 first-order features and 15 texture features are screened out from the 107 imaging features by the SelectKBest function combined with chi-square test or analysis of variance, and are used as the ROI 1 imaging features for discriminating the existence of pancreatic surgical area inflammation.

[0029] As a preferred scheme of the present application, the method for constructing the inflammation existence discrimination model comprises:

[0030] Four imaging models are constructed by taking the ROI 1 imaging features as input and the inflammation existence result as output through logistic regression, support vector machine, decision tree and random forest;

[0031] The performances of the four imaging models are comprehensively evaluated by using sensitivity, specificity, accuracy, precision, F1 score and area under the receiver operating characteristic curve;

[0032] The model with the best comprehensive performance is selected from the four imaging models, and the ROI 1 imaging features and clinical risk factors are taken as combined input and the inflammation existence result is taken as output to construct the inflammation existence discrimination model.

[0033] As a preferred scheme of the present application, the method for taking the ROI 1 imaging features and clinical risk factors as combined input comprises:

[0034] The ROI 1 imaging features and clinical risk factors are assigned weights through cross-attention to obtain the attention weights of the ROI 1 imaging features and clinical risk factors, wherein:

[0035] The attention weight of the ROI 1 imaging features , wherein, is the attention weight vector of the ROI 1 imaging features, is a query Query vector from the clinical risk factors, is a key Key vector from the ROI 1 imaging features, is a value Value vector from the ROI 1 imaging features, is the vector dimension of

[0036] The attention weight of the clinical risk factors , wherein, is the attention weight vector of the clinical risk factors, is a query Query vector from the ROI 1 imaging features, is a key Key vector from the clinical risk factors, a value vector Value from clinical risk factors, a vector dimension of ;

[0037] combining the ROI 1 radiomics features and the attention weights of the clinical risk factors with the ROI 1 radiomics features and the clinical risk factors to obtain ROI 1 radiomics enhanced features and clinical risk factors enhanced features ;

[0038] evaluating the difference between the ROI 1 radiomics enhanced features and the clinical risk factors enhanced features and the ROI 1 radiomics features and the clinical risk factors, and correcting the ROI 1 radiomics enhanced features and the clinical risk factors enhanced features according to the difference to obtain ROI 1 radiomics comprehensive features and clinical risk factors comprehensive features ;

[0039] combining the ROI 1 radiomics comprehensive features and the clinical risk factors comprehensive features to form the combination input;

[0040] the combination input is: ;

[0041] wherein, is the combination input, is the ROI 1 radiomics comprehensive features, is the clinical risk factors comprehensive features, is a channel concatenation function, is the ROI 1 radiomics features, is the clinical risk factors features, is the ROI 1 radiomics enhanced features, is the clinical risk factors enhanced features, is the attention weight of , is the attention weight of , is a similarity function.

[0042] As a preferred scheme of the present application, the clinical risk factors include lesion location, lesion size and presence or absence of main pancreatic duct dilation.

[0043] As a preferred scheme of the present application, the inflammation presence result includes presence of inflammation and absence of inflammation.

[0044] As a preferred scheme of the present application, the inflammations existence discrimination model is subjected to explainability analysis by using the SHAP framework.

[0045] As a preferred scheme of the present application, the inflammations existence discrimination model is constructed by using a random forest.

[0046] Compared with the prior art, the present application has the following beneficial effects:

[0047] In the fusion model constructed by combining the risk factors and the imaging features, the LPD operation area connective tissue ROI 1 is segmented by using the TotalSegmentator segmentation model and the nnUNet segmentation framework, and the ROI 1 imaging features are extracted, the clinical risk factors are found by using the retrospective study, and the screening operation support is provided, so as to ensure the inflammations existence discrimination accuracy of the fusion model, and the clinical risk factors and the imaging features are subjected to attention mechanism splicing processing, the key features are highlighted, the key features are quickly captured by using the fusion model, and the inflammations existence discrimination efficiency of the fusion model is ensured. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are only exemplary, and other drawings can be obtained by the provided drawings without creative labor for those skilled in the art.

[0049] Figure 1 The pancreatic perivascular fat image feature based operation area inflammations existence discrimination method flowchart provided for the embodiments of the present application;

[0050] Figure 2 The operation area inflammations existence discrimination model construction flowchart provided for the embodiments of the present application;

[0051] Figure 3 The surgery video screenshot of the pancreatic mesenteric area connective tissue inflammation condition provided for the embodiments of the present application;

[0052] Figure 4 The perioperative outcome data chart of the patients in the non-inflammation group and the inflammation group provided for the embodiments of the present application;

[0053] Figure 5 The single factor analysis result chart provided for the embodiments of the present application;

[0054] Figure 6 The multi-factor binary logistic regression analysis result chart provided for the embodiments of the present application;

[0055] Figure 7A patient screening and grouping chart provided for an embodiment of the present application;

[0056] Figure 8 A cross-region imageomics feature schematic diagram provided for an embodiment of the present application;

[0057] Figure 9 A 4-imageomics model performance evaluation data chart provided for an embodiment of the present application;

[0058] Figure 10 A 4-imageomics model performance radar chart provided for an embodiment of the present application;

[0059] Figure 11 An inflammation presence discrimination model performance evaluation data chart provided for an embodiment of the present application;

[0060] Figure 12 An inflammation presence discrimination model performance evaluation curve chart provided for an embodiment of the present application

[0061] Figure 13 An inflammation presence discrimination model SHAP analysis result chart provided for an embodiment of the present application;

[0062] Figure 14 An ROI 1 segmentation result chart provided for an embodiment of the present application. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0064] As shown in Figure 1 and Figure 2 , the present application provides a pancreatic surgical area inflammation presence discrimination method fusing fat-related image features, comprising the following steps:

[0065] Determine the clinical risk factors for pancreatic surgical area inflammation presence discrimination by statistical methods;

[0066] On CT venous phase images, segment the LPD (Laparoscopic Pancreaticoduodenectomy) surgical area connective tissue ROI 1 by the TotalSegmentator segmentation model and the nnUNet segmentation framework, and extract the ROI 1 imageomics features for pancreatic surgical area inflammation presence discrimination on the LPD surgical area connective tissue ROI 1.

[0067] By means of various machine learning algorithms, an inflammation existence discrimination model for discriminating the existence of pancreatic surgical area inflammation is constructed based on the ROI 1 imageomics features and clinical risk factors.

[0068] The present application constructs an inflammation existence discrimination model for discriminating the existence of pancreatic surgical area inflammation by combining clinical risk factors and imageomics features, that is, fuses multi-modal features in the discrimination of the existence of pancreatic surgical area inflammation, provides diversified information for the discrimination of the existence of pancreatic surgical area inflammation, and realizes the improvement of discrimination accuracy.

[0069] In order to ensure the discrimination accuracy of the inflammation existence discrimination model, the single factor analysis in the statistical method is used to find the differences of the LPD patients in different inflammation groups in the perioperative outcomes and clinical data, the clinical data showing statistically significant differences between groups in the single factor analysis is included in the multi-factor binary logistic regression analysis, to identify the independent risk factors of inflammation of pancreatic mesenteric connective tissue (i.e. existence of inflammation), so as to objectively select the lesion location, lesion size and presence or absence of main pancreatic duct dilation as clinical risk factors, to provide clinical features in the discrimination of the existence of inflammation, to objectively screen out the most relevant clinical features of the inflammation of pancreatic mesenteric connective tissue in the clinical data, and to improve the discrimination accuracy of the inflammation existence discrimination model in using clinical features to discriminate the existence of inflammation.

[0070] Similarly, the application uses the combination method of TotalSegmentator segmentation model (from Wasserthal, J., Breit, H.-C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D., Cyriac, J., Yang, S., Bach, M., Segeroth, M., 2023. TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images. Radiology: Artificial Intelligence. https: / / doi.org / 10.1148 / ryai.230024.) and nnUNet segmentation framework on CT venous phase images to segment the LPD operation area connective tissue ROI 1, which can realize high-precision segmentation of the LPD operation area connective tissue ROI 1. High-precision segmentation of the LPD operation area connective tissue ROI 1 corresponds to high-precision image feature extraction. That is, in the high-precision segmented LPD operation area connective tissue ROI 1, the ROI 1 image feature that guarantees high-performance discrimination of inflammation existence can be obtained, which realizes objective screening of the image feature most related to inflammation in the pancreatic mesenteric connective tissue in clinical data, and improves the discrimination accuracy of the inflammation existence discrimination model using image features to discriminate inflammation existence.

[0071] After determining the clinical risk factors and ROI 1 image features, the application constructs four image feature models for inflammation existence discrimination on the ROI 1 image features through four machine learning algorithms, namely logistic regression (Logistic Regression, LR), support vector machine (Support Vector Machine, SVM), decision tree (Decision Tree, DT) and random forest (Random Forest, RF), and compares the performance of the four image feature models. The best model is selected as the ROI 1 image feature model for pancreatic operation area inflammation existence discrimination.

[0072] Then, by combining the ROI 1 image feature and the clinical risk factor as a combined input on the ROI 1 image feature model, a clinical-image feature fusion model, that is, an inflammation existence discrimination model for pancreatic operation area inflammation existence discrimination, is constructed.

[0073] In order to improve the operation efficiency of the inflammation existence discrimination model, the cross attention mechanism is adopted to quickly capture key information in the combined input obtained by combining the ROI 1 imageomics features and the clinical risk factors, so as to capture the correlation between the ROI 1 imageomics features and the clinical risk factors, realize the information interaction between the two modalities, and assign attention weights to the two modalities according to the importance of the interaction information in the process of realizing the information interaction between the two modalities, so as to give high weights to the features in the ROI 1 imageomics features that have high correlation with the clinical risk factors, and give high weights to the features in the clinical risk factors that have high correlation with the ROI 1 imageomics features, so as to highlight the key information related to the clinical risk factors in the ROI 1 imageomics features, and highlight the key information related to the ROI 1 imageomics features in the clinical risk factors, so as to realize the interaction of multi-modal features, effectively establish complementary links between different modalities, help the model to identify and utilize the correlation between different modalities, and improve the efficiency performance of the inflammation existence discrimination model.

[0074] Further, in order to avoid the multi-modal interaction between the ROI 1 imageomics features and the clinical risk factors due to the cross attention mechanism, so that the key information related to the clinical risk factors in the ROI 1 imageomics features is too prominent, and the key information related to the ROI 1 imageomics features in the clinical risk factors is too prominent, so that the model can only focus on the interaction information, and the remaining information in the clinical risk factors and the ROI 1 imageomics features except the interaction information will be ignored due to the small assigned weight, and there is also information beneficial to the inflammation existence discrimination in these remaining information, that is, it will cause the loss of feature information in the discrimination process, therefore, the present application limits the degree of multi-modal interaction, that is, on the basis of the ROI 1 imageomics features (i.e. ROI 1 imageomics enhanced features) obtained by highlighting the key information related to the clinical risk factors through cross attention weight, the original ROI 1 imageomics features are reintroduced, so as to supplement the lost feature information of the ROI 1 imageomics features after the cross attention enhancement, and obtain the ROI 1 imageomics comprehensive features, and on the basis of the clinical risk factors (i.e. clinical risk factor enhanced features) obtained by highlighting the key information related to the ROI 1 imageomics features through cross attention weight, the original clinical risk factors are reintroduced, so as to supplement the lost feature information of the clinical risk factors after the cross attention enhancement, and obtain the clinical risk factor comprehensive features.

[0075] In the process of supplementing the lost feature information of the ROI 1 radiomics features due to cross attention enhancement, the supplement intensity is associated with the similarity of the ROI 1 radiomics enhanced features and the original ROI 1 radiomics features, the lower the similarity of the ROI 1 radiomics enhanced features and the original ROI 1 radiomics features, the greater the cross attention enhancement degree, the more the lost feature information, and the more the lost feature information that needs to be supplemented, so the higher the supplement intensity should be given to the original ROI 1 radiomics features, the higher the weight given to the original ROI 1 radiomics features in the present application to correspond to the supplement intensity, that is, the lower the similarity of the ROI 1 radiomics enhanced features and the original ROI 1 radiomics features, the higher the weight given to the original ROI 1 radiomics features to supplement a large amount of lost feature information, the higher the similarity of the ROI 1 radiomics enhanced features and the original ROI 1 radiomics features, the lower the weight given to the original ROI 1 radiomics features to supplement a small amount of lost feature information.

[0076] Similarly, in the process of supplementing the lost feature information of the clinical risk factors due to cross attention enhancement, the supplement intensity is associated with the similarity of the clinical risk factor enhanced features and the original clinical risk factor features, the lower the similarity of the clinical risk factor enhanced features and the original clinical risk factor features, the greater the cross attention enhancement degree, the more the lost feature information, and the more the lost feature information that needs to be supplemented, so the higher the supplement intensity should be given to the clinical risk factors, the higher the weight given to the original clinical risk factors in the present application to correspond to the supplement intensity, that is, the lower the similarity of the clinical risk factor enhanced features and the original clinical risk factors, the higher the weight given to the original clinical risk factors to supplement a large amount of lost feature information, the higher the similarity of the clinical risk factor enhanced features and the original clinical risk factors, the lower the weight given to the original clinical risk factors to supplement a small amount of lost feature information.

[0077] The determination method of the clinical risk factors comprises:

[0078] The basic clinical data, laboratory examination data, preoperative abdominal enhanced CT data, pathological data and perioperative outcome data are acquired, wherein:

[0079] Patient grouping: a pancreatic surgeon who has passed the LPD learning curve reviews the operation video of the patient, divides the patient into a non-inflammation group, a mild inflammation group and a severe inflammation group according to the inflammation of the connective tissue in the pancreatic mesentery area between the pancreatic head and the SMV-PV axis (superior mesenteric vein-portal vein axis), records the SMV-PV axis dissection time and specimen resection time, and saves the operation video screenshot (such as Figure 3The surgical video screenshots of different inflammation conditions of the pancreatic mesenteric region connective tissue were shown, a was no inflammation; b was mild inflammation; c was severe inflammation. Another pancreatic surgeon who passed the LPD learning curve reviewed the surgical video screenshots again and recorded the patient inflammation group. When the two people were inconsistent in the inflammation group of the same patient, the patient group was finally determined through discussion. In addition, the patient's pathological results were reviewed, and the patients who mistakenly identified the tumor invasion of SMV-PV, GDA, SMA, pancreatic uncinate process mesentery or pancreatic head rear vascular groove as inflammatory adhesion were excluded. The SMV-PV axis dissection time refers to the time required to disconnect the pancreatic neck to completely strip the connective tissue in front of and right behind the SMVPV axis. The specimen resection time refers to the time required to implant the trocar sleeve needle to completely remove the tissues and organs involved in the LPD, including the pancreatic head, distal stomach, duodenum, partial jejunum, bile duct and gallbladder.

[0080] Basic clinical data: gender, age, body mass index (BMI), history of weight loss, ASA classification, history of abdominal pain, history of acute pancreatitis (AP) / chronic pancreatitis (CP), history of neoadjuvant chemotherapy, history of biliary drainage.

[0081] Laboratory examination data: the first test results after the patient's admission were collected. White blood cell count, monocyte count and C-reactive protein (CRP) reflect the patient's inflammatory condition; hemoglobin and albumin reflect the patient's preoperative nutritional status; direct bilirubin (DBil), γ-glutamyl transpeptidase (γ-GT) and alkaline phosphatase (ALP) reflect biliary obstruction; blood amylase (AMY) and blood lipase (LPS) reflect pancreatic injury.

[0082] Preoperative abdominal enhanced CT data: examination interval (interval between CT examination date and operation date), lesion length, lesion location, main pancreatic duct dilation, bile duct dilation.

[0083] Pathological data: lesion nature, tumor microvessel infiltration.

[0084] Perioperative outcome data: surgical procedure, SMV-PV dissection time, specimen resection time, operation time, intraoperative blood transfusion, clinically relevant postoperative pancreatic fistula (CR-POPF), bile fistula, delayed gastric emptying (DGE), post-pancreatectomy hemorrhage (PPH), infection, postoperative complication classification (Clavien-Dindo classification system), postoperative hospital stay, and in-hospital mortality.

[0085] like Figure 4 The perioperative outcomes of patients in the non-inflammatory and inflammatory groups are shown in the figure. The median SMV-PV axis dissection time, specimen resection time, and operation time were 27.8 minutes, 108.7 minutes, and 250 minutes in the non-inflammatory group, respectively, while they were 44.7 minutes, 153.4 minutes, and 300 minutes in the inflammatory group, respectively. The Mann-Whitney U test showed statistically significant differences in all these values ​​(Z=-7.848,-8.054,-5.626; P<0.001). All 17 patients undergoing combined SMV-PV repair and 4 patients converting to open surgery had inflammation of the pancreatic mesangial connective tissue, and the difference in surgical approach was statistically significant (Fisher's exact test, P<0.001). Intraoperative blood transfusion occurred in 7 patients (6.25%) in the non-inflammatory group and 15 patients (19.74%) in the inflammatory group, with a statistically significant difference (χ²). 2 (1) = 5.441, P = 0.020). In the non-inflammatory group, 12 patients (10.7%) developed Clavien-Dindo grade III or higher complications, compared to 24 patients (26.4%) in the inflammatory group; the difference was statistically significant (χ²). 2 (1) = 8.439, P = 0.004). In the non-inflammatory group, 35 patients (28.6%) had microvascular invasion, while in the inflammatory group, 49 patients (53.8%) had it, a statistically significant difference (χ²). 2 (1) = 10.568, P = 0.001). In addition, there were no statistically significant differences in short-term outcome indicators such as clinically relevant pancreatic fistula, bile fistula, delayed gastric emptying, postoperative bleeding, infection, postoperative hospital stay and outcome.

[0086] For quantitative data in basic clinical data, laboratory test data, preoperative enhanced abdominal CT data, pathological data, and perioperative outcome data, the Shapiro-Wilk test was used to determine whether they conformed to a normal distribution. Among them:

[0087] For continuous variables in continuous quantitative data that conform to a normal distribution, the mean (Mean) and standard deviation (SD) are used to describe them (Mean±SD). When the variances of two groups are homogeneous, the independent samples t-test is used; otherwise, the Mann-Whitney U test is used.

[0088] For variables in the continuous data that do not conform to a normal distribution, the median (M), the 25th percentile (Q1), and the 75th percentile (Q3) were used to describe the data [M(Q1-Q3)]. The Mann-Whitney U test was used for comparisons between groups.

[0089] Basic clinical data, laboratory test data, preoperative enhanced abdominal CT data, pathological data, and perioperative outcome data were described as counts and percentages. Chi-square tests were used for comparisons between groups.

[0090] When comparing two groups, Fisher's exact test is used when there is a theoretical frequency, and the continuous correction chi-square test is used when there is a frequency of 1 < theoretical frequency < 5.

[0091] There is no correction formula for comparisons between multiple groups. Fisher's exact test is used when the number of cells with a theoretical frequency < 1 or a number with a theoretical frequency < 5 exceeds 20% of the total number of cells.

[0092] like Figure 5 and Figure 6 As shown, stepwise regression was used to screen variables with statistical differences in univariate analysis, and multivariate binary logistic regression analysis was used to identify clinical risk factors.

[0093] The median age of patients in the non-inflammatory group was 58 (50-65) years, and the median BMI was 22.4 (20.5-24.2) kg / m², of which 66 were male (58.9%). The median age of patients in the inflammatory group was 60 (54-64) years, and the median BMI was 22.5 (20.5-24.6) kg / m², of which 55 were male (60.4%). The Mann-Whitney U test showed no statistically significant differences in age and BMI (Z=-0.988, -0.223; P values ​​were 0.323 and 0.823, respectively), and the chi-square test showed no statistically significant difference in sex (χ²). 2(1)=0.048, P=0.827). Univariate analysis showed that the differences in history of weight loss, abdominal pain, pancreatitis, neoadjuvant chemotherapy and biliary drainage between the no-inflammation and inflammation groups were not statistically significant (P values were 0.058, 0.311, 0.814, 0.448 and 0.619, respectively). In addition, no statistically significant differences in laboratory data such as white blood cell count, monocyte count, CRP, hemoglobin, albumin, direct bilirubin, gamma glutamyl transpeptidase, alkaline phosphatase, blood amylase and blood lipase were observed between the two groups.

[0094] The median interval time between CT examinations in the no-inflammation and inflammation groups was 8.0 (4.0-12.0) days and 7.5 (5.0-13.3) days, respectively, and Mann-Whitney U test showed that the difference was not statistically significant (Z=-0.188, P=0.851). The median lesion length reported by CT imaging in the no-inflammation and inflammation groups was 2.0 (1.4-3.3) cm and 2.6 (2.0-3.4) cm, respectively, and Mann-Whitney U test showed that the difference was statistically significant (Z=-2.224, P=0.025). After classification of the CT lesion length according to the AJCC liver, biliary and pancreatic tumor T stage

[40] , the number of patients with CT lesion length ≤2 cm in the no-inflammation and inflammation groups was 44 (53.7%) and 26 (32.9%), respectively, and chi-square test showed that the difference was statistically significant (χ 2 (1)=7.722, P=0.005); the number of patients with lesion length of 2-4 cm was 26 (31.7%) and 47 (59.5%), respectively, and chi-square test showed that the difference was statistically significant (χ 2 (1)=7.722, P<0.001); the number of patients with lesion length >4 cm was 12 (14.6%) and 6 (7.6%), respectively, and chi-square test showed that the difference was not statistically significant (χ 2 (1)=1.548, P=0.213). The number of patients with CT imaging reported lesion located in the pancreatic head in the no-inflammation and inflammation groups was 41 (38.0%) and 61 (70.1%), respectively, and chi-square test showed that the difference was statistically significant (χ 2 (1)=19.967, P<0.001); the number of patients with lesion in the ampulla was 34 (31.5%) and 12 (13.8%), respectively, and chi-square test showed that the difference was statistically significant (χ 2 (1)=8.364, P=0.004); the number of patients with lesion in the common bile duct was 21 (19.4%) and 11 (12.7%), respectively, and chi-square test showed that the difference was not statistically significant (χ 2(1) = 1.625, P = 0.202); duodenal lesions were found in 12 cases (11.1%) and 3 cases (3.4%), respectively. The chi-square test showed that the difference was statistically significant (χ²). 2 (1) = 3.985, P = 0.046). The number of patients with main pancreatic duct dilation reported on CT images in the non-inflammatory group was 46 (41.4%), while in the inflammatory group it was 59 (64.8%). The chi-square test showed a statistically significant difference (χ²). 2 (1) = 10.964, P = 0.001). Furthermore, the CT imaging reports showed bile duct dilatation in 65 cases (58.6%) in the non-inflammatory group and 52 cases (57.1%) in the inflammatory group, with no statistically significant difference (χ²). 2 (1) = 0.041, P = 0.839).

[0095] Univariate analysis identified statistically significant preoperative data as independent variables, with the occurrence of pancreatic mesangial connective tissue inflammation as the dependent variable (no inflammation = 0; inflammation = 1). Multivariate binary logistic regression analysis was then performed. The independent variables included in the multivariate analysis were: the extent of main pancreatic duct dilation reported on CT imaging, the long diameter of the lesion (≤2cm or >2cm), and the location of the lesion (ampullary lesion, pancreatic head lesion, and duodenal lesion). Multivariate analysis showed that dilation of the main pancreatic duct reported on CT images (OR=2.580, 95%CI: 1.234-5.395, P=0.012) was a risk factor for inflammation of the connective tissue in the pancreatic mesangial region, while lesion length ≤2cm (OR=0.405, 95%CI: 0.202-0.813, P=0.011) and ampullary lesions (OR=0.315, 95%CI: 0.116-0.856, P=0.023) were protective factors against inflammation of the connective tissue in the pancreatic mesangial region. Multivariate analysis showed that CT-reported pancreatic head lesions (OR=1.087, 95%CI: 0.605-1.950, P=0.781) and duodenal lesions (OR=0.166, 95%CI: 0.020-1.380, P=0.096) were independent risk factors for inflammation of the non-pancreatic mesangial connective tissue.

[0096] The segmentation methods for the connective tissue ROI 1 of the LPD surgical area include:

[0097] CT venous phase images were preprocessed using window width and level adjustment and nonlinear filtering.

[0098] On the preprocessed CT venous phase images, the important anatomical structures during LPD were segmented using the TotalSegmentator segmentation model, and clipping frames including the pancreas, SMV-PV axis, celiac trunk, common hepatic artery and its main branches were determined.

[0099] Segmenting the abdominal visceral fat in the local region by the nnUNet segmentation framework on the CT venous phase image in the clipping frame;

[0100] Segmenting the abdominal visceral fat in the global region by the nnUnet segmentation framework on the pre-processed CT venous phase image;

[0101] Spatially registering the local segmentation result of the abdominal visceral fat with the global segmentation result of the abdominal visceral fat and performing mask operation to obtain the LPD surgical area connective tissue ROI 1.

[0102] In the present application, the nnUNet segmentation framework is used to segment the abdominal visceral tissue in the clipping frame containing the pancreas, SMV-PV axis, celiac trunk, common hepatic artery and its main branch structure. This is a local image segmentation based on CT images, which realizes fine segmentation focusing on the key area. However, compared with global image segmentation, local image segmentation may lose part of the image information, resulting in a loss of segmentation accuracy. In order to make up for the loss of segmentation accuracy caused by the loss of information, the nnUNet segmentation framework is also used to segment the abdominal visceral tissue on the original CT image, which is a global image segmentation based on CT images. The global image segmentation result is added to the local image segmentation result through spatial registration and mask operation, which is equivalent to adding the original image information to the local segmentation to reduce information loss and avoid the loss of segmentation accuracy caused by information loss. This realizes further improvement of the segmentation accuracy of the hierarchical modeling strategy, as follows:

[0103] The method for spatially registering the local segmentation result of the abdominal visceral fat with the global segmentation result of the abdominal visceral fat and performing mask operation includes:

[0104] Spatially register the global segmentation result of the abdominal visceral fat with the clipping frame to obtain the global segmentation result of the abdominal visceral fat in the clipping frame coordinate system;

[0105] Extract the global segmentation result of the abdominal visceral fat in the clipping frame from the global segmentation result of the abdominal visceral fat in the clipping frame coordinate system;

[0106] Real-time obtain the segmentation accuracy performance of the nnUNet segmentation framework for local segmentation of the abdominal visceral fat, and real-time obtain the segmentation accuracy performance of the nnUNet segmentation framework for global segmentation of the abdominal visceral fat;

[0107] Perform mask operation on the global segmentation result of the abdominal visceral fat in the clipping frame and the local segmentation result of the abdominal visceral fat in a weighted fusion manner according to the segmentation accuracy of the local segmentation and the segmentation accuracy of the global segmentation to obtain the LPD surgical area connective tissue ROI 1;

[0108] The expression of the mask operation is:

[0109] ;

[0110] In the formula, is the connective tissue in the LPD surgery area, is the global segmentation result of the abdominal visceral fat in the cropping frame, is the local segmentation result of the abdominal visceral fat, is the segmentation accuracy of the global segmentation, is the segmentation accuracy of the local segmentation.

[0111] In the mask operation, the weighted fusion method is adopted, the real-time segmentation accuracy of the local segmentation and the global segmentation is introduced into the nnUNet segmentation framework, that is, the performance of the global segmentation model and the performance of the local segmentation model of the nnUNet segmentation framework at the current time are judged, the higher the performance is, the higher the segmentation result confidence is, therefore, after normalization, the weight of the global segmentation result and the local segmentation result can add the segmentation precision of the local segmentation loss through the global segmentation, so as to achieve the purpose of "adding the global image segmentation result to the local image segmentation result, which is equivalent to adding the original image information in the local segmentation to reduce the information loss, so as to avoid the segmentation precision loss caused by the information loss, and further improve the segmentation precision of the hierarchical modeling strategy".

[0112] The segmentation gold standard of ROI 1 to ROI 6 is obtained by manual segmentation by a radiologist with professional experience in abdominal imaging.

[0113] In the present application, the real-time segmentation accuracy of the local segmentation and the global segmentation of the nnUNet segmentation framework is measured by the training set obtained by collecting case samples, that is, the error between the local segmentation result and the global segmentation result obtained by the nnUNet segmentation framework and the segmentation gold standard of ROI 1 obtained by manual segmentation by a radiologist with professional experience in abdominal imaging in the training set is calculated, the higher the error is, the lower the segmentation performance is, and the lower the confidence of the segmentation result is.

[0114] The segmentation accuracy is quantified by the difference between the global segmentation result of the abdominal visceral fat in the cropping frame and the local segmentation result of the abdominal visceral fat and the segmentation gold standard of ROI 1, and the expression of the segmentation accuracy quantification is:

[0115] ;

[0116] ;

[0117] In the formula, the global segmentation result of the nnUNet segmentation framework for the current time instant on the i-th image sample in the cropping frame, the local segmentation result of the nnUNet segmentation framework for the current time instant on the i-th image sample, the segmentation gold standard of ROI 1 on the i-th image sample, is the mean square loss operation formula, m is the total number of image samples, which is consistent with the number of training set cases.

[0118] The local segmentation result of the abdominal visceral fat is spatially registered with the global segmentation result of the abdominal visceral fat, and a mask operation is performed to obtain the LPD surgical area connective tissue ROI 1 (marked as Original when predicting the presence of inflammation, as shown in Figs. 14a-c).

[0119] The method for extracting the ROI 1 imageomic features includes:

[0120] 107 imageomic features are extracted from the LPD surgical area connective tissue ROI 1 by the PyRadiomics component in the Python library;

[0121] The SelectKBest function is combined with the chi-square test or analysis of variance to screen 20 important imageomic features consisting of 1 shape feature, 4 first-order features and 15 texture features from the 107 imageomic features, as shown in Figure 8 , and used as the ROI 1 imageomic features for discriminating the presence of pancreatic surgical area inflammation.

[0122] When predicting the presence of inflammation in the pancreatic mesenteric area connective tissue of LPD patients, 142 patients enter the training set, of which 78 are non-inflammation patients and 64 are inflammation patients; 61 patients enter the validation set, of which 34 are non-inflammation patients and 27 are inflammation patients. When predicting the degree of inflammation, 63 patients enter the training set, of which 30 are mild inflammation patients and 33 are severe inflammation patients; 28 patients enter the validation set, of which 14 are mild inflammation patients and 14 are severe inflammation patients. In addition, there are 16 patients in the external validation set formed by the prospective cohort, of which 8 are non-inflammation patients, 4 are mild inflammation patients, and 4 are severe inflammation patients. The screening and grouping of patients are shown in Figure 7 .

[0123] The method for constructing the inflammation presence discrimination model includes:

[0124] Four radiomics models were constructed by using Logistic Regression (LR), Support Vector Machine (SVM), Decision Tree (DT) and Random Forest (RF) respectively, with the radiomics features of ROI1 as input and the presence of inflammation as output.

[0125] The performance of the four radiomics models was evaluated comprehensively by using Sensitivity (SEN), Specificity (SPE), Accuracy (ACC), Precision (PRE), F1 score and Area Under the Curve (AUC) of Receiver Operating Characteristic Curve (ROC).

[0126] In predicting the presence or absence of inflammation, based on the 20 important radiomics features extracted from ROI1, four machine learning algorithms, LR, SVM, DT and RF, were used to construct radiomics models (Rad-score1), i.e. radiomics models. As shown in FIG. 1, the sensitivity, specificity, accuracy, precision, F1 score and AUC of the radiomics models constructed by the four machine learning algorithms in the training set and the validation set are shown. Figure 9 Figure 10 Figure 10 is a radar chart drawn according to the above data. The closer the position of the performance parameter in the radar chart to the outermost circle, the better the model performs in that parameter; the larger the area covered by the radar chart, the stronger the comprehensive performance of the model. In the training set, Figure 9 and Figure 10 Among the four radiomics models, the RF model was superior to the other three models in terms of performance parameters except that the sensitivity was slightly lower than that of DT, and its AUC was 0.982 (95% CI: 0.963-0.994) and the accuracy was 0.937. In the validation set, the RF model was superior to the other three models in terms of performance parameters except that the sensitivity was lower than that of the other three machine learning algorithms, and its AUC was 0.807 (95% CI: 0.691-0.913) and the accuracy was 0.770; the prediction performance of the DT model was inferior to that of the RF, and its AUC was 0.752 (95% CI: 0.637-0.856) and the accuracy was 0.689. Among the four machine learning algorithms, the performance of LR and SVM was relatively poor. Therefore, the RF algorithm was finally selected to construct the radiomics model for predicting the presence or absence of inflammation (i.e. the model with the best comprehensive performance among the four radiomics models).​​

[0127] In the four image-based models, the model with the best comprehensive performance was selected as the inflammation presence discrimination model, which was trained with the ROI 1 image-based features and the clinical risk factors as the combined input and the inflammation presence result as the output.

[0128] When predicting the presence or absence of inflammation, the performance of the clinical model (a RF model trained with the clinical risk factors as the input and the inflammation presence result as the output), the best image-based model (i.e., the model with the best comprehensive performance in the four image-based models), and the fusion model (i.e., the inflammation presence discrimination model) was compared as shown in Figure 11 and Figure 12 In the validation set and the external validation set, the fusion model had the best performance, with the AUC being 0.826 and 0.734, respectively, and the accuracy being 0.770 and 0.688, respectively. The image-based model ranked second, with the AUC being 0.807 and 0.766, respectively, and the accuracy being 0.770 and 0.625, respectively. However, the difference in the AUC between the fusion model and the image-based model was not statistically significant (P values were 0.360 and 0.400, respectively, as shown by the DeLong test). The clinical model had the worst performance in predicting the inflammation of the pancreatic mesenteric connective tissue in the training set, the validation set, and the external validation set, with the AUC being 0.634, 0.678, and 0.539, respectively, and the difference in the AUC between the clinical model and the fusion model was statistically significant (P values were both less than 0.001, as shown by the DeLong test). In the validation set, the decision curve analysis showed that the fusion model would produce greater clinical net benefits when the threshold probability was between 0.18 and 0.84, and the calibration curve of the fusion model showed that the predicted results and the actual situation of the presence or absence of inflammation had good consistency, as shown in Figure 12 .

[0129] The method of using the ROI 1 image-based features and the clinical risk factors as the combined input includes the following steps:

[0130] The ROI 1 image-based features and the clinical risk factors are assigned weights through cross-attention to obtain the attention weights of the ROI 1 image-based features and the clinical risk factors, wherein:

[0131] The attention weight of the ROI 1 image-based features , wherein, is the attention weight vector of the ROI 1 image-based features, is a query vector from the clinical risk factors, is a key vector from the ROI 1 image-based features, and a value vector Value from the ROI 1 radiomics features, a vector dimension of ;

[0132] attention weights of clinical risk factors , wherein, is an attention weight vector of the clinical risk factors, is a query vector Query from the ROI 1 radiomics features, is a key vector Key from the clinical risk factors, is a value vector Value from the clinical risk factors, a vector dimension of ;

[0133] combining the attention weights of the ROI 1 radiomics features and the clinical risk factors with the ROI 1 radiomics features and the clinical risk factors to obtain ROI 1 radiomics enhanced features and clinical risk factor enhanced features ;

[0134] In order to improve the operation efficiency of the inflammation existence discrimination model, the present application quickly captures key information in the combined input obtained by combining the ROI 1 radiomics features and the clinical risk factors, adopts a cross-attention mechanism, can capture the correlation between the two modal features of the ROI 1 radiomics features and the clinical risk factors, realize the information interaction between the two modal features, and at the same time, according to the importance of the interaction information, the attention weights of the two modal features are allocated, the high weight is given to the features in the ROI 1 radiomics features which have high correlation with the clinical risk factors, and the high weight is given to the features in the clinical risk factors which have high correlation with the ROI 1 radiomics features, so as to highlight the key information related to the clinical risk factors in the ROI 1 radiomics features, and highlight the key information related to the ROI 1 radiomics features in the clinical risk factors, so as to realize the interaction of multi-modal features, effectively establish complementary relationship between different modalities, help the model to identify and utilize the correlation between different modalities, and improve the efficiency performance of the inflammation existence discrimination model.

[0135] evaluate the difference between the ROI 1 radiomics enhanced features and the clinical risk factor enhanced features and the ROI 1 radiomics features and the clinical risk factors, and correct the ROI 1 radiomics enhanced features and the clinical risk factor enhanced features generalization to obtain ROI 1 radiomics comprehensive features and clinical risk factor comprehensive features ;

[0136] Furthermore, this invention aims to avoid the multimodal interaction between ROI 1 radiomics features and clinical risk factors caused by the cross-attention mechanism, which would result in either overemphasis on key information related to clinical risk factors in ROI 1 radiomics features or overemphasis on key information related to ROI 1 radiomics features in clinical risk factors. This would cause the model to focus only on the interaction information, while the remaining information in clinical risk factors and ROI 1 radiomics features, excluding the interaction information, would be ignored due to their small weight allocation. This remaining information also contains information that is helpful in determining the presence of inflammation, which would lead to the loss of feature information during the discrimination process.

[0137] Therefore, this invention limits the degree of multimodal interaction. Specifically, based on the ROI 1 radiomics features obtained by highlighting key information related to clinical risk factors through cross-attention weights (i.e., ROI 1 radiomics enhanced features), the original ROI 1 radiomics features are reintroduced. This replenishes the feature information lost by the ROI 1 radiomics features after cross-attention enhancement, resulting in comprehensive ROI 1 radiomics features. Similarly, based on the clinical risk factors obtained by highlighting key information related to ROI 1 radiomics features through cross-attention weights (i.e., clinical risk factor enhanced features), the original clinical risk factors are reintroduced. This replenishes the feature information lost by the clinical risk factors after cross-attention enhancement, resulting in comprehensive clinical risk factor features.

[0138] Specifically, in the process of supplementing the feature information lost by cross-attention enhancement in ROI 1 radiomics features, the strength of the supplementation is related to the similarity between the enhanced ROI 1 radiomics features and the original ROI 1 radiomics features. Correlation, similarity between ROI 1 radiomics enhanced features and original ROI 1 radiomics features The lower the value, the greater the degree of cross-attention enhancement, the more feature information is lost, and the more lost feature information needs to be supplemented. Therefore, the original ROI 1 radiomics features should be assigned a higher supplementation strength. This invention assigns a higher weight to the original ROI 1 radiomics features to correspond to the supplementation strength, that is, when the similarity between the enhanced ROI 1 radiomics features and the original ROI 1 radiomics features is high... The lower the similarity, the higher the weight assigned to the original ROI 1 radiomics features to compensate for the large amount of lost feature information. This is because the similarity between the ROI 1 radiomics enhancement features and the original ROI 1 radiomics features is... The higher the value, the lower the weight assigned to the original ROI 1 radiomics features to compensate for the small amount of lost feature information.

[0139] Similarly, in the process of supplementing the feature information lost due to the cross-attention enhancement of the clinical risk factor, the supplement intensity is similar to the similarity between the enhanced feature of the clinical risk factor and the original feature of the clinical risk factor Correlation, the similarity between the enhanced feature of the clinical risk factor and the original feature of the clinical risk factor The lower, the greater the cross-attention enhancement degree, the more the lost feature information, and the more the lost feature information that needs to be supplemented, so the higher the supplement intensity should be given to the clinical risk factor, and the higher the weight given to the original clinical risk factor in the present application to correspond to the supplement intensity, that is, the similarity between the enhanced feature of the clinical risk factor and the original feature of the clinical risk factor The lower, the higher the weight given to the original clinical risk factor to supplement a large amount of lost feature information, the similarity between the enhanced feature of the clinical risk factor and the original feature of the clinical risk factor The higher, the lower the weight given to the original clinical risk factor to supplement a small amount of lost feature information.

[0140] The ROI 1 imageomics comprehensive feature And the clinical risk factor comprehensive feature Combination to form a combination input;

[0141] The combination input is: ;

[0142] In the formula, The combination input, The ROI 1 imageomics comprehensive feature, The clinical risk factor comprehensive feature, The channel splicing function, The ROI 1 imageomics feature, The clinical risk factor feature, The ROI 1 imageomics enhanced feature, The clinical risk factor enhanced feature, The attention weight of The attention weight of The similarity function. The clinical risk factor includes lesion location, lesion size, and presence or absence of main pancreatic duct dilation.

[0143] The inflammation presence result includes inflammation and no inflammation.

[0144] The inflammation presence result includes inflammation and no inflammation.

[0145] As Figure 13 ​As shown, the SHAP (Shapley Additive exPlanations) framework is used to perform explainability analysis on the inflammation presence discrimination model.

[0146] The extracted imageomics feature types are limited to first-order features, shape features and texture features, which are well associated with the pathological mechanism of inflammation. First-order features reflect the gray distribution characteristics of a single voxel (or pixel), which can capture the homogeneity changes of tissue density, including 90th percentile, mean and range, etc. Shape features reflect the appearance and size of ROI, which can quantify the distortion of anatomical structures, such as minor axis length, shape elongation and maximum 3D diameter, etc. Texture features can analyze the heterogeneity of tissues, including the following five categories: gray level co-occurrence matrix (GLCM), gray level dependence matrix (GLDM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), and neighborhood gray tone difference matrix (NGTDM). In the inflammation presence discrimination model, the top 20 imageomics features with the largest contribution to the model are shown in the Shapley Additive exPlanations (SHAP) values Figure 13 As shown, when the imageomics feature SHAP value is positive, it increases the model prediction probability, and when it is negative, it reduces the model prediction probability.

[0147] The inflammation presence discrimination model is constructed by a random forest (RF).

[0148] In the fusion model constructed by combining risk factors and imageomics features, the LPD surgical area connective tissue ROI 1 is segmented by the TotalSegmentator segmentation model and the nnUNet segmentation framework, and the ROI 1 imageomics features are extracted. The clinical risk factors are found by retrospective study, which supports the screening operation, thereby ensuring the inflammation presence discrimination accuracy of the fusion model. At the same time, the clinical risk factors and imageomics features are processed by attention mechanism splicing, which highlights the key features, so that the fusion model can quickly capture the key features, thereby ensuring the inflammation presence discrimination efficiency of the fusion model.

[0149] The above embodiments are only exemplary embodiments of the present application, and are not intended to limit the present application. The scope of protection of the present application is defined by the claims. Those skilled in the art can make various modifications or equivalent replacements to the present application within the spirit and protection scope of the present application, and such modifications or equivalent replacements should also be considered to fall within the protection scope of the present application.

Claims

1. A method of discriminating the presence of inflammation in a pancreatic surgical field by fusing fat-related image features, characterized by, The method comprises the following steps: determining clinical risk factors for discriminating the presence of inflammation in the pancreatic surgical area by statistical methods; segmenting the LPD surgical area connective tissue ROI 1 by combining the TotalSegmentator segmentation model and the nnUNet segmentation framework on the CT venous phase image, and extracting ROI 1 image features for discriminating the presence of inflammation in the pancreatic surgical area on the LPD surgical area connective tissue ROI 1; constructing an inflammation presence discrimination model for discriminating the presence of inflammation in the pancreatic surgical area based on the ROI 1 image features and the clinical risk factors by using multiple machine learning algorithms; the inflammation presence discrimination model takes the ROI 1 image features and the clinical risk factors as combined inputs, and takes the inflammation presence result as output; the ROI 1 image features and the clinical risk factors are assigned weights by cross-attention to obtain the attention weights of the ROI 1 image features and the clinical risk factors; The attention weight of the ROI 1 imaging omics feature and the clinical risk factor is combined with the ROI 1 imaging omics feature and the clinical risk factor to obtain an ROI 1 imaging omics enhanced feature and the clinical risk factor enhanced feature ; evaluate the difference between the ROI 1 radiomics enhanced features and the clinical risk factors, and correct the ROI 1 radiomics enhanced features according to the difference and the generalization of the clinical risk factors enhanced features to obtain the ROI 1 radiomics comprehensive features and the clinical risk factors comprehensive features ; The ROI 1 imageomics comprehensive feature and clinical risk factor comprehensive feature Combination constitutes the combination input ; In the formula, is a channel concatenation function, is an ROI 1 imageomics feature, is a clinical risk factor feature, is is an attention weight of is is an attention weight of is a similarity function. 2.The method of claim 1, wherein the method comprises: obtaining a plurality of image features of the pancreatic surgery area from a plurality of images of the pancreatic surgery area; and fusing the image features to obtain a fused image feature. the determination method of the clinical risk factors comprises: obtaining basic clinical data, laboratory examination data, preoperative abdominal enhanced CT data, pathological data, and perioperative outcome data; for the measurement data in the basic clinical data, laboratory examination data, preoperative abdominal enhanced CT data, pathological data, and perioperative outcome data, Shapiro-Wilk test is used to determine whether it conforms to the normal distribution, wherein: for continuous variables in the measurement data that conform to the normal distribution, the mean and standard deviation are used for description, and two independent sample t test is used for comparison when the variances of the two groups are equal, otherwise Mann-Whitney U test is used; for variables in the measurement data that do not conform to the non-normal distribution, the median, 25th percentile and 75th percentile are used for description, and Mann-Whitney U test is used for comparison between groups; the count data in the basic clinical data, laboratory examination data, preoperative abdominal enhanced CT data, pathological data, and perioperative outcome data are described by the number of examples and percentage, and chi-square test is used for comparison between groups, wherein: for comparison between two groups, Fisher's exact test is used when there is a theoretical frequency, and continuous correction chi-square test is used when there is 1 When there is no correction formula for comparison among multiple groups, Fisher's exact test is used when the theoretical frequency is less than 1 or the number of cells with 1 Stepwise regression is used to screen variables with statistical differences in single factor analysis, and multivariate binary logistic regression analysis is used to determine the clinical risk factors. 3.The method of claim 1, wherein the method comprises: obtaining a plurality of image features of the pancreatic surgery area from a plurality of images of the pancreatic surgery area; and fusing the image features to obtain a fused image feature. The segmentation method of the LPD surgical area connective tissue ROI 1 comprises: preprocessing the CT venous phase image by window width and window level adjustment and nonlinear filtering processing; segmenting the LPD intraoperative important anatomical structure on the preprocessed CT venous phase image by the TotalSegmentator segmentation model to determine a clipping frame containing the pancreas, SMV-PV axis, celiac trunk, common hepatic artery and its main branch structure; segmenting the abdominal visceral fat locally on the CT venous phase image in the clipping frame by the nnUNet segmentation framework; On the pretreated CT venous phase image, visceral fat in the abdomen is globally segmented by the nnUnet segmentation framework; The local segmentation result of the abdominal visceral fat is spatially registered with the global segmentation result of the abdominal visceral fat, and a mask operation is performed to obtain the LPD surgical area connective tissue ROI 1.

4. The method of claim 3, wherein the method is a method of discriminating the presence of pancreatic surgical field inflammation by fusing fat-related image features. The method for extracting the ROI 1 imageomics features comprises: 107 imageomics features are extracted from the LPD surgical area connective tissue ROI 1 by the PyRadiomics component in the Python library; 20 important imageomics features consisting of 1 shape feature, 4 first-order features and 15 texture features are screened out from the 107 imageomics features by the SelectKBest function combined with chi-square test or analysis of variance, and are used as the ROI 1 imageomics features for discriminating the existence of pancreatic surgical area inflammation.

5. The method of claim 4, wherein the method is a method of discriminating the presence of pancreatic surgical field inflammation by fusing fat-related image features. The method for constructing the inflammation existence discrimination model comprises: Four imageomics models are constructed by taking the ROI 1 imageomics features as input and the inflammation existence result as output through logistic regression, support vector machine, decision tree and random forest; The performances of the four imageomics models are comprehensively evaluated by using sensitivity, specificity, accuracy, precision, F1 score and area under the receiver operating characteristic curve; The model with the best comprehensive performance is selected from the four imageomics models, and the inflammation existence discrimination model is constructed by taking the ROI 1 imageomics features and the clinical risk factors as combined input and the inflammation existence result as output.

6. The pancreatic surgical area inflammation existence discrimination method fusing fat-related image features according to claim 1, wherein: the attention weight of the ROI 1 radiomics features , wherein, is an attention weight vector of the ROI 1 radiomics features, is a query vector from clinical risk factors, is a key vector from the ROI 1 radiomics features, is a value vector from the ROI 1 radiomics features, is a vector dimension of attention weight of the clinical risk factor , wherein, is an attention weight vector of the clinical risk factor, is a query vector from the ROI 1 radiomics features, is a key vector from the clinical risk factors, is a value vector from the clinical risk factors, is a vector dimension of .

7. The method of claim 2, wherein the method is a method of discriminating the presence of pancreatic surgical field inflammation by fusing fat-related image features, characterized by: The clinical risk factors include lesion location, lesion size and presence or absence of main pancreatic duct dilation. 8.The method of claim 1, wherein the method comprises: obtaining a plurality of images of a pancreas of a patient; and determining a presence of inflammation in a pancreatic surgery area of the patient based on the plurality of images. The inflammation existence result includes inflammation and no inflammation.

9. The method of claim 1, wherein the method is a method of discriminating the presence of pancreatic surgical field inflammation by fusing fat-related image features. The inflammation existence discrimination model is subjected to explainability analysis by using the SHAP framework.

10. The method of claim 1, wherein the method is a method of discriminating the presence of pancreatic surgical field inflammation by fusing fat-related image features. The inflammation existence discrimination model is constructed by random forest.

Images