Pancreatic operation area inflammation existence discrimination method fusing fat related image features

Abstract

The invention relates to the technical field of medical imaging omics analysis, in particular to a pancreatic operation area inflammation existence judgment method fused with fat related image features, which comprises the following steps of: determining clinical risk factors through a statistical method; the method comprises the following steps: on a CT vein phase image, segmenting a connective tissue ROI (Region of Interest) 1 in an LPD (Local Parkinson Distortion) operation area through combination of a TotalSegmentor segmentation model and an nnUNet segmentation framework, and extracting image omics characteristics of the ROI 1 on the connective tissue ROI 1 in the LPD operation area; and through a plurality of machine learning algorithms, based on the ROI 1 radiomics features and clinical risk factors, constructing an inflammation existence discrimination model. In the fusion model constructed by combining the risk factors and the radiomics features, retrospective research finds that the clinical risk factors have screening operation support, and meanwhile, the clinical risk factors and the radiomics features are subjected to splicing processing by adopting an attention mechanism, so that the fusion model quickly captures key features, and the fusion accuracy is improved. Therefore, the inflammatory existence judgment precision and efficiency of the fusion model are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of medical imaging omics analysis, and particularly to a method for discriminating the presence of inflammation in a pancreatic surgical area by integrating fat-related imaging features. Background Art

[0002] Pancreaticoduodenectomy (PD) is the standard surgical procedure for treating lesions of the pancreatic head and periampullary region. Resection involves the dense vascular and neural plexus within the mesopancreatic region (e.g., SMA, SMV / PV, etc.), and therefore is considered one of the most complex and challenging general surgical procedures. Inflammation in the mesopancreatic region (e.g., chronic pancreatitis, fibrosis after neoadjuvant therapy) replaces the loose perivascular spaces with dense fibrous tissue, leading to a loss of normal perivascular anatomical layers and a significant increase in the incidence of "difficult PD."

[0003] Preoperative contrast-enhanced CT is the primary method for assessing inflammation in the surgical area, but technical limitations limit the clinical value of conventional CT findings. Although CT images clearly demonstrate the anatomy of the large vessels surrounding the pancreas, their low soft tissue resolution can prevent the identification of early microscopic inflammatory features (such as microcirculatory impairment), leading to underestimation of the extent of intraoperative inflammation in the pancreatic mesentery. Furthermore, existing imaging assessment strategies often focus on morphological changes in the "tumor-vascular contact area," overlooking the impact of inflammation on the mechanical properties of local tissues, a key factor in predicting the difficulty of vascular dissection.

[0004] Radiomics extracts texture, morphology, and functional features from medical images at high throughput, transforming microscopic tissue heterogeneity (such as the spatial distribution of inflammation-related fibrosis) into computable high-dimensional data. This data has the potential to be used to construct clinically meaningful inflammation prediction models. Therefore, radiomics currently holds high clinical value in determining the presence of inflammation in surgical areas. It is even combined with clinical risk factors to construct fusion models that fuse multimodal features for discriminating the presence of inflammation. However, the selection of clinical risk factors and radiomics features in fusion model construction is often based on subjective expert experience, lacking objective screening computational support. This impacts the accuracy of the fusion model in discriminating the presence of inflammation. Furthermore, clinical risk factors and radiomics features are often directly concatenated, making it difficult for the fusion model to quickly capture key features, thus affecting its efficiency in discriminating the presence of inflammation. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for discriminating the presence of inflammation in the pancreatic surgical area by integrating fat-related imaging features, so as to solve the technical problems in the existing technology that the screening of clinical risk factors and imaging features in the clinical-radiological fusion model is usually based on the subjective experience of experts, which affects the accuracy of the fusion model in discriminating the presence of inflammation, and clinical risk factors and imaging features are usually directly spliced, which makes it difficult for the fusion model to quickly capture key features, affecting the efficiency of the fusion model in discriminating the presence of inflammation.

[0006] In order to solve the above technical problems, the present invention specifically provides the following technical solutions: A method for determining the presence of inflammation in the pancreatic surgical area by integrating fat-related imaging features comprises the following steps: Clinical risk factors for the presence of pancreatic surgical site inflammation were determined by statistical methods; On CT venous phase images, the TotalSegmentator segmentation model and the nnUNet segmentation framework were combined to segment the LPD surgical area connective tissue ROI 1. Radiomic features of ROI 1 for discriminating the presence of pancreatic inflammation were extracted from the LPD surgical area connective tissue ROI 1. Through multiple machine learning algorithms, an inflammation presence discrimination model for discriminating the presence of inflammation in the pancreatic surgical area was constructed based on the ROI 1 imaging features and clinical risk factors.

[0007] As a preferred embodiment of the present invention, the method for determining clinical risk factors includes: Basic clinical data, laboratory test data, preoperative abdominal enhanced CT data, pathological data, and perioperative outcome data were obtained; The Shapiro-Wilk test was used to determine whether the measurement data in basic clinical data, laboratory test data, preoperative abdominal enhanced CT data, pathological data, and perioperative outcome data conformed to normal distribution. The continuous variables in the measurement data that conformed to the normal distribution were described by means of mean and standard deviation. The two independent sample t-test was used for comparison between the two groups when the variances were equal; otherwise, the Mann-Whitney U test was used. The variables that did not conform to the normal distribution in the measurement data were described using the median, 25th percentile, and 75th percentile, and the Mann-Whitney U test was used for comparison between groups; The enumeration data of basic clinical data, laboratory examination data, preoperative abdominal enhanced CT data, pathological data, and perioperative outcome data were described as number of cases and percentages. The chi-square test was used for comparison between groups. For comparisons between the two groups, Fisher's exact test was used when theoretical frequencies existed, and the continuity-corrected chi-square test was used when 1 < theoretical frequency < 5 existed; When multiple groups were compared without a correction formula, Fisher's exact test was used when the theoretical frequency was <1 or when the number of cells with 1 ≤ theoretical frequency <5 exceeded 20% of the total number of cells; Stepwise regression was used to screen the variables with statistical significance in univariate analysis, and multivariate binary logistic regression analysis was used to identify clinical risk factors.

[0008] As a preferred embodiment of the present invention, the segmentation method of the LPD area connective tissue ROI 1 includes: The CT venous phase images were preprocessed by adjusting the window width and window position and performing nonlinear filtering. On the pre-processed CT venous phase images, the TotalSegmentator segmentation model was used to segment the important anatomical structures during LPD surgery, and a cropping frame was determined that included the pancreas, SMV-PV axis, celiac trunk, common hepatic artery and its main branches. On the CT venous phase images within the cropping box, the abdominal visceral fat is locally segmented using the nnUNet segmentation framework; On the pre-processed CT venous phase images, global segmentation of abdominal visceral fat was performed using the nnUnet segmentation framework; The local segmentation result of the abdominal visceral fat is spatially registered and masked with the global segmentation result of the abdominal visceral fat to obtain the LPD operation area connective tissue ROI 1.

[0009] As a preferred embodiment of the present invention, the method for extracting radiomics features of ROI 1 includes: 107 radiomics features were extracted from the connective tissue ROI 1 of the LPD surgical area using the PyRadiomics component in the Python library; Twenty important imaging features consisting of one shape feature, four first-order features, and 15 texture features were screened out from 107 imaging features using the SelectKBest function combined with the chi-square test or analysis of variance. These features were used as the ROI 1 imaging features for discriminating the presence of inflammation in the pancreatic surgical area.

[0010] As a preferred embodiment of the present invention, the method for constructing the inflammation presence discrimination model includes: Four radiomics models were constructed using logistic regression, support vector machine, decision tree, and random forest, all with ROI 1 radiomics features as input and inflammation presence results as output; The performance of the four radiomics models was comprehensively evaluated using sensitivity, specificity, accuracy, precision, F1 score, and area under the receiver operating characteristic curve (AUC). The model with the best comprehensive performance was selected from the four radiomics models. The ROI 1 radiomics features and clinical risk factors were used as combined inputs, and the inflammation presence results were used as output to construct the inflammation presence discrimination model.

[0011] As a preferred embodiment of the present invention, the method using ROI 1 radiomics features and clinical risk factors as combined input includes: The attention weights of ROI 1 radiomic features and clinical risk factors were obtained by cross-attention allocation, where: Attention weights of the radiomics features of ROI 1 , where is the attention weight vector of ROI 1 radiomics features, is the query vector derived from clinical risk factors, The key vector from the radiomics features of ROI 1, is the value vector from the radiomics feature of ROI 1, for The vector dimension of Attention weighting of the clinical risk factors , where is the attention weight vector of clinical risk factors, is the query vector from the radiomics features of ROI 1, Key vectors derived from clinical risk factors, is the value vector from clinical risk factors, for The vector dimension of The attention weights of ROI 1 radiomics features and clinical risk factors were combined with the ROI 1 radiomics features and clinical risk factors to obtain the ROI 1 radiomics enhanced features. and clinical risk factor enhancement features ; Assess the differences between ROI 1 radiomics enhancement features and clinical risk factor enhancement features and ROI 1 radiomics features and clinical risk factors, and modify ROI 1 radiomics enhancement features based on the differences and clinical risk factor enhancement features Generalization, obtaining comprehensive radiomics features of ROI 1 and comprehensive characteristics of clinical risk factors ; The radiomics comprehensive features of ROI 1 and comprehensive characteristics of clinical risk factors Combining to form said combined input; The combined input is: ; Where, For combined input, is the comprehensive radiomics feature of ROI 1, is a comprehensive feature of clinical risk factors. is the channel splicing function, is the radiomics feature of ROI 1, Clinical risk factor characteristics, For ROI 1 radiomics enhancement features, Enhanced characterization of clinical risk factors, for The attention weight, for The attention weight, is the similarity function.

[0012] As a preferred embodiment of the present invention, the clinical risk factors include lesion location, lesion size, and the presence or absence of main pancreatic duct dilatation.

[0013] As a preferred embodiment of the present invention, the inflammation presence results include inflammation and no inflammation.

[0014] As a preferred solution of the present invention, the SHAP framework is used to perform interpretability analysis on the inflammation presence discrimination model.

[0015] As a preferred embodiment of the present invention, the inflammation presence discrimination model is constructed by random forest.

[0016] Compared with the prior art, the present invention has the following beneficial effects: In the fusion model constructed by combining risk factors and imaging omics features, the present invention uses the TotalSegmentator segmentation model and the nnUNet segmentation framework to segment the connective tissue ROI 1 of the LPD surgical area and extract the imaging omics features of ROI 1. Retrospective studies are used to discover clinical risk factors, and screening operations are supported to ensure the accuracy of the fusion model in distinguishing the presence of inflammation. At the same time, clinical risk factors and imaging omics features are spliced ​​using the attention mechanism to highlight key features, allowing the fusion model to quickly capture key features, thereby ensuring the efficiency of the fusion model in distinguishing the presence of inflammation. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other implementation drawings based on the provided drawings without inventive effort.

[0018] Figure 1 A flow chart of a method for determining the presence of inflammation in the surgical area based on pancreatic perivascular fat imaging features provided by an embodiment of the present invention; Figure 2 A flowchart of a model for discriminating the presence of inflammation in the surgical area provided in an embodiment of the present invention; Figure 3 Screenshot of a surgical video recording of inflammation of the connective tissue in the mesangial region of the pancreas provided in an embodiment of the present invention; Figure 4 Perioperative outcome data of patients in the non-inflammation group and the inflammation group provided in the embodiment of the present invention; Figure 5 A graph showing the results of a single factor analysis according to an embodiment of the present invention; Figure 6 A graph showing the results of a multi-factor binary logistic regression analysis provided by an embodiment of the present invention; Figure 7 A diagram showing the screening and grouping of patients provided in an embodiment of the present invention; Figure 8 A schematic diagram of cross-regional imaging omics features provided by an embodiment of the present invention; Figure 9 Performance evaluation data of four radiomics models provided in the embodiments of the present invention; Figure 10 Radar charts showing the performance of four radiomics models provided by the embodiments of the present invention; Figure 11 A performance evaluation data diagram of the inflammation presence discrimination model provided in an embodiment of the present invention; Figure 12 Performance evaluation curve diagram of the inflammation presence discrimination model provided by the embodiment of the present invention Figure 13 This is a diagram showing the SHAP analysis results of the inflammation presence discrimination model provided by an embodiment of the present invention; Figure 14 This is the segmentation result diagram of ROI 1 provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0020] like Figure 1 and Figure 2 As shown, the present invention provides a method for determining the presence of inflammation in the pancreatic surgical area by integrating fat-related imaging features, comprising the following steps: Clinical risk factors for the presence of pancreatic surgical site inflammation were determined by statistical methods; On CT venous phase images, the TotalSegmentator segmentation model and the nnUNet segmentation framework were combined to segment the connective tissue ROI 1 of the LPD (laparoscopic pancreaticoduodenectomy) surgical area. Radiomic features for identifying inflammation in the pancreatic surgical area were extracted from the connective tissue ROI 1 of the LPD surgical area. Through multiple machine learning algorithms, an inflammation presence discrimination model for pancreatic surgical area inflammation was constructed based on ROI 1 radiomics features and clinical risk factors.

[0021] The present invention constructs an inflammation presence discrimination model for discriminating the presence of inflammation in the pancreatic surgical area by combining clinical risk factors and imaging genomics features. That is, multimodal features are integrated in the discrimination of the presence of inflammation in the pancreatic surgical area to provide diversified information for the discrimination of the presence of inflammation in the pancreatic surgical area, thereby improving the discrimination accuracy.

[0022] To ensure the discriminant accuracy of the inflammation presence discriminant model, the present invention utilizes univariate analysis within the statistical method to identify differences in perioperative outcomes and clinical data among LPD patients in different inflammation groups. Clinical data showing statistically significant between-group differences in univariate analysis were then incorporated into a multivariate binary logistic regression analysis to identify independent risk factors for the presence of inflammation in the pancreatic mesentery connective tissue (i.e., the presence of inflammation). Lesion location, lesion size, and the presence of main pancreatic duct dilatation were objectively selected as clinical risk factors to provide clinical features for discriminating the presence of inflammation. This objectively screens the clinical features most relevant to the presence of inflammation in the pancreatic mesentery connective tissue from the clinical data, thereby improving the accuracy of the inflammation presence discriminant model in using clinical features to discriminate the presence of inflammation.

[0023] Similarly, the present invention uses a combination of the TotalSegmentator segmentation model (derived from Wasserthal, J., Breit, H.-C., Meyer, MT, Pradella, M., Hinck, D., Sauter, AW, Heye, T., Boll, D., Cyriac, J., Yang, S., Bach, M., Segeroth, M., 2023. TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CTImages. Radiology: Artificial Intelligence. https: / / doi.org / 10.1148 / ryai.230024.) and the nnUNet segmentation framework on CT venous phase images to segment the LPD surgical area connective tissue ROI 1, which can achieve high-precision segmentation of the LPD surgical area connective tissue ROI 1. The high-precision segmentation of 1 corresponds to the extraction of high-precision imaging omics features. That is to say, the imaging omics features of ROI 1 that guarantee high-performance discrimination of the presence of inflammation can be obtained in the connective tissue ROI 1 of the LPD surgical area obtained by high-precision segmentation. This enables the objective screening of the imaging omics features that are most relevant to whether the connective tissue in the pancreatic mesentery is inflamed in the clinical data, thereby improving the accuracy of the inflammation discrimination model in using imaging omics features to discriminate the presence of inflammation.

[0024] After determining clinical risk factors and ROI 1 radiomics features, the present invention constructed four radiomics models for discriminating the presence of inflammation based on the ROI 1 radiomics features using four machine learning algorithms: logistic regression (LR), support vector machine (SVM), decision tree (DT), and random forest (RF). The performance of these four radiomics models was compared, and the model with the best overall performance was selected as the ROI 1 radiomics model for discriminating the presence of inflammation in the pancreatic surgical area. Then, by combining the ROI 1 radiomics features and clinical risk factors as combined inputs in the ROI 1 radiomics model, a fusion model of clinical-radiological features was constructed, that is, an inflammation presence discrimination model for discriminating the presence of inflammation in the pancreatic surgical area.

[0025] In order to improve the computational efficiency of the inflammation presence discrimination model, the present invention quickly captures key information in the combined input obtained by combining ROI 1 radiomics features and clinical risk factors, and adopts a cross-attention mechanism. This can capture the correlation between the two modal features between ROI 1 radiomics features and clinical risk factors, and realize information interaction between the two modal features. At the same time, in the process of realizing information interaction between the two modal features, attention weights are assigned to the two modal features according to the importance of the interactive information, so as to give high weights to features in ROI 1 radiomics features that have high correlation with clinical risk factors, and give high weights to features in clinical risk factors that have high correlation with ROI 1 radiomics features, thereby highlighting the key information related to clinical risk factors in ROI 1 radiomics features, and highlighting the key information related to ROI 1 radiomics features in clinical risk factors. This achieves the interaction of multimodal features, effectively establishes complementary connections between different modalities, helps the model identify and utilize the correlation between different modalities, and improves the efficiency and performance of the inflammation presence discrimination model.

[0026] Furthermore, in order to avoid the multimodal interaction between ROI 1 radiomics features and clinical risk factors caused by the cross-attention mechanism, the present invention causes the key information related to clinical risk factors in the ROI 1 radiomics features to be overly prominent, and the key information related to the ROI 1 radiomics features in the clinical risk factors to be overly prominent, so that the model can only focus on the interactive information, and the remaining information in the clinical risk factors and ROI 1 radiomics features except the interactive information will be ignored due to the small weight assigned. Among these remaining information, there is also information that is beneficial to the discrimination of the presence of inflammation, that is, it will cause the loss of feature information in the discrimination process. Therefore, the present invention limits the degree of multimodal interaction, that is, on the basis of the ROI 1 radiomics features (i.e., ROI 1 radiomics enhanced features) obtained by highlighting the key information related to the clinical risk factors through cross-attention weights, the original ROI 1 radiomics features are reintroduced, thereby supplementing the feature information lost after the ROI 1 radiomics features are enhanced through cross-attention, and obtaining the ROI 1 radiomics comprehensive features, as well as highlighting the key information related to the ROI 1 through cross-attention weights. 1. Based on the clinical risk factors obtained from the key information related to the radiomics features (i.e., clinical risk factor enhancement features), the original clinical risk factors were reintroduced, thereby supplementing the characteristic information lost after the clinical risk factors were enhanced by cross-attention, and obtaining the comprehensive features of clinical risk factors.

[0027] Among them, in the process of supplementing the feature information lost after the ROI 1 imaging omics feature is enhanced due to cross-attention, the supplement intensity is associated with the similarity between the ROI 1 imaging omics enhancement feature and the original ROI 1 imaging omics feature. The lower the similarity between the ROI 1 imaging omics enhancement feature and the original ROI 1 imaging omics feature, 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, a higher supplement intensity should be given to the original ROI1 imaging omics feature. The present invention gives a higher weight to the original ROI 1 imaging omics feature to correspond to the supplement intensity, that is, when the similarity between the ROI 1 imaging omics enhancement feature and the original ROI 1 imaging omics feature is lower, a higher weight is given to the original ROI 1 imaging omics feature to supplement a large amount of lost feature information. When the similarity between the ROI 1 imaging omics enhancement feature and the original ROI 1 imaging omics feature is higher, a lower weight is given to the original ROI 1 imaging omics feature to supplement a small amount of lost feature information.

[0028] Similarly, in the process of supplementing the characteristic information lost due to cross-attention enhancement of clinical risk factors, the intensity of supplementation is associated with the similarity between the enhanced characteristics of clinical risk factors and the characteristics of original clinical risk factors. The lower the similarity between the enhanced characteristics of clinical risk factors and the characteristics of original clinical risk factors, the greater the degree of cross-attention enhancement, the more characteristic information is lost, and the more lost characteristic information needs to be supplemented. Therefore, a higher supplementation intensity should be given to clinical risk factors. The present invention gives a higher weight to the original clinical risk factors to correspond to the supplementation intensity, that is, when the similarity between the enhanced characteristics of clinical risk factors and the original clinical risk factors is lower, a higher weight is given to the original clinical risk factors to supplement a large amount of lost characteristic information. When the similarity between the enhanced characteristics of clinical risk factors and the original clinical risk factors is higher, a lower weight is given to the original clinical risk factors to supplement a small amount of lost characteristic information.

[0029] Clinical risk factors can be determined by: Basic clinical data, laboratory test data, preoperative abdominal enhanced CT data, pathological data, and perioperative outcome data were obtained, including: Patient grouping: A pancreatic surgeon who had gone through the LPD learning curve reviewed the patients' surgical video recordings and divided them into the non-inflammation group, the mild inflammation group, and the severe inflammation group based on the inflammation of the pancreatic mesentery connective tissue between the pancreatic head and the SMV-PV axis (superior mesenteric vein-portal vein axis). The time of SMV-PV axis dissection and specimen removal was recorded, and screenshots of the surgical video recordings that clearly showed the inflammation of the pancreatic mesentery connective tissue were saved (e.g. Figure 3(See figure) Screenshots from surgical videos showing different levels of inflammation in the pancreatic mesentery (a: no inflammation; b: mild inflammation; c: severe inflammation). A second pancreatic surgeon who had completed the learning curve for LPD reviewed the surgical video screenshots and recorded the patient's inflammation grouping. When the two surgeons disagreed on the inflammation grouping for the same patient, the patient's grouping was finalized through discussion. Furthermore, pathological findings were reviewed to exclude patients in whom tumor invasion of the SMV-PV, GDA, SMA, pancreatic uncinate mesentery, or the posterior pancreatic head vascular groove was misidentified as inflammatory adhesions. The SMV-PV axis dissection time refers to the time required from severing the pancreatic neck to completely dissecting the connective tissue anterior and right posterior to the SMVPV axis. The specimen resection time refers to the time required from trocar insertion to complete resection of the tissues and organs involved in LPD (including the pancreatic head, distal stomach, duodenum, part of the jejunum, bile duct, and gallbladder).

[0030] Basic clinical data included gender, age, body mass index (BMI), history of emaciation, ASA classification, history of abdominal pain, history of acute pancreatitis (AP) / chronic pancreatitis (CP), history of neoadjuvant chemotherapy, and history of biliary drainage.

[0031] Laboratory test data: The first test results after admission were collected. These included white blood cell count, monocyte count, and C-reactive protein (CRP) to reflect inflammation; hemoglobin and albumin to reflect preoperative nutritional status; direct bilirubin (DBil), γ-glutamyl transpeptidase (γ-GT), and alkaline phosphatase (ALP) to reflect biliary obstruction; and serum pancreatic amylase (AMY) and lipase (LPS) to reflect pancreatic injury.

[0032] Preoperative abdominal enhanced CT data: examination interval (the interval between CT examination date and surgery date), lesion long diameter, lesion location, main pancreatic duct dilatation, and bile duct dilatation.

[0033] Pathological data: nature of lesions, tumor microvascular infiltration.

[0034] Perioperative outcome data included surgical method, SMV-PV dissection time, specimen resection time, operation time, intraoperative blood transfusion, clinically relevant postoperative pancreatic fistula (CR-POPF), biliary fistula, delayed gastric emptying (DGE), postpancreatectomy hemorrhage (PPH), infection, postoperative complication grade (Clavien-Dindo grade), postoperative hospital stay, and in-hospital death.

[0035] like Figure 4 Perioperative outcome data for patients in the non-inflammation group and the inflammation group are shown in the figure. The median time for SMV-PV axis dissection, specimen removal, and operation was 27.8 minutes, 108.7 minutes, and 250 minutes in the non-inflammation group, and 44.7 minutes, 153.4 minutes, and 300 minutes in the inflammation group, respectively. The Mann-Whitney U test showed that the differences were statistically significant (Z = -7.848, -8.054, -5.626; all P < 0.001). Inflammation of the pancreatic mesentery connective tissue was observed in 17 patients undergoing combined SMV-PV repair and 4 patients undergoing conversion to laparotomy. The difference between surgical approaches was statistically significant (Fisher's exact test, P < 0.001). Intraoperative blood transfusion occurred in 7 patients (6.25%) in the non-inflammation group and 15 patients (19.74%) in the inflammation group, with a statistically significant difference (χ2). 2 (1) = 5.441, P = 0.020). Twelve patients (10.7%) in the non-inflammation group developed complications of Clavien-Dindo grade III or higher, while 24 patients (26.4%) in the inflammation group developed complications. The difference was statistically significant (χ 2 (1) = 8.439, P = 0.004). In the non-inflammation group, 35 patients (28.6%) had microvascular infiltration, while in the inflammation group, 49 patients (53.8%) had microvascular infiltration, and the difference was statistically significant (χ 2 (1)=10.568, P=0.001). In addition, there were no statistically significant differences in clinically relevant short-term outcome indicators such as pancreatic fistula, biliary fistula, delayed gastric emptying, postoperative bleeding, infection, postoperative hospital stay and prognosis.

[0036] The Shapiro-Wilk test was used to determine whether the measurement data in basic clinical data, laboratory test data, preoperative abdominal enhanced CT data, pathological data, and perioperative outcome data conformed to normal distribution. Continuous variables that conformed to normal distribution in the measurement data were described using mean ± SD. When the variances between the two groups were equal, the independent sample t test was used; otherwise, the Mann-Whitney U test was used. For variables that did not conform to the normal distribution in the measurement data, the median (M), the 25th percentile (First Quartile, Q1), and the 75th percentile (Third Quartile, Q3) were used to describe [ M (Q1-Q3)], and the Mann-Whitney U test was used for comparison between groups; The enumeration data of basic clinical data, laboratory examination data, preoperative abdominal enhanced CT data, pathological data, and perioperative outcome data were described as number of cases and percentages. The chi-square test was used for comparison between groups. For comparisons between the two groups, Fisher's exact test was used when theoretical frequencies existed, and the continuity-corrected chi-square test was used when 1 < theoretical frequency < 5 existed; When multiple groups were compared without a correction formula, Fisher's exact test was used when the theoretical frequency was <1 or when the number of cells with 1 ≤ theoretical frequency <5 exceeded 20% of the total number of cells; like Figure 5 and Figure 6 As shown, stepwise regression method was used to screen variables with statistical differences in univariate analysis, and multivariate binary logistic regression analysis was used to determine clinical risk factors.

[0037] The median age of patients in the non-inflammation group was 58 (50-65) years, and the median BMI was 22.4 (20.5-24.2) kg / m2, of which 66 patients (58.9%) were male. The median age of patients in the inflammation group was 60 (54-64) years, and the median BMI was 22.5 (20.5-24.6) kg / m2, of which 55 patients (60.4%) were male. The Mann-Whitney U test showed that there was no statistically significant difference in age and BMI (Z=-0.988, -0.223; P values ​​were 0.323 and 0.823, respectively). The chi-square test showed that there was no statistically significant difference in gender (χ 2(1) = 0.048, P = 0.827). The results of univariate analysis showed that there were no statistically significant differences in the history of weight loss, abdominal pain, pancreatitis, neoadjuvant chemotherapy, and biliary drainage between the non-inflammation group and the inflammation group (P values ​​were 0.058, 0.311, 0.814, 0.448, and 0.619, respectively). In addition, no statistically significant differences were observed in laboratory test data such as white blood cell count, monocyte count, CRP, hemoglobin, albumin, direct bilirubin, γ-glutamyl transpeptidase, alkaline phosphatase, blood amylase, and blood lipase between the two groups.

[0038] The median intervals between CT examinations in the non-inflammation group and the inflammation group were 8.0 (4.0-12.0) days and 7.5 (5.0-13.3) days, respectively. The Mann-Whitney U test showed that there was no statistically significant difference (Z=-0.188, P=0.851). The median long diameter of the lesions reported by CT imaging in the non-inflammation group and the inflammation group were 2.0 (1.4-3.3) cm and 2.6 (2.0-3.4) cm, respectively. The Mann-Whitney U test showed that there was a statistically significant difference (Z=-2.224, P=0.025). After classifying the long diameter of CT lesions with reference to the AJCC hepatobiliary and pancreatic tumor T staging

[40] , the number of patients with lesions with a long diameter of ≤2 cm reported by CT imaging in the non-inflammation group and the inflammation group were 44 (53.7%) and 26 (32.9%), respectively. The chi-square test showed that there was a statistically significant difference (χ 2 (1) = 7.722, P = 0.005); the lesion length was 2-4 cm in 26 patients (31.7%) and 47 patients (59.5%), respectively. The chi-square test showed that the difference was statistically significant (χ 2 (1) = 7.722, P < 0.001); there were 12 patients (14.6%) and 6 patients (7.6%) with lesion length > 4 cm, respectively. The chi-square test showed no statistically significant difference (χ 2 (1)=1.548, P=0.213). The number of patients with CT imaging reports of lesions located in the pancreatic head in the non-inflammation group and the inflammation group was 41 (38.0%) and 61 (70.1%), respectively. The chi-square test showed that the difference was statistically significant (χ 2 (1) = 19.967, P < 0.001); there were 34 cases (31.5%) and 12 cases (13.8%) of ampulla lesions, respectively. The chi-square test showed that the difference was statistically significant (χ 2 (1) = 8.364, P = 0.004); the common bile duct lesions were 21 cases (19.4%) and 11 cases (12.7%), respectively. The chi-square test showed that there was no statistically significant difference (χ 2(1) = 1.625, P = 0.202); duodenal lesions were 12 cases (11.1%) and 3 cases (3.4%), respectively. The chi-square test showed that the difference was statistically significant (χ 2 (1) = 3.985, P = 0.046). In the non-inflammation group, 46 patients (41.4%) were found to have dilatation of the main pancreatic duct on CT imaging, while in the inflammation group, 59 patients (64.8%) were found to have dilatation of the main pancreatic duct. The chi-square test showed that the difference was statistically significant (χ 2 (1) = 10.964, P = 0.001). In addition, the CT imaging reported bile duct dilatation in 65 cases (58.6%) and 52 cases (57.1%) in the non-inflammation group and the inflammation group, respectively, with no statistically significant difference (χ 2 (1)=0.041,P=0.839).

[0039] Preoperative data that showed statistically significant differences in univariate analysis were used as independent variables, and multivariate binary logistic regression analysis was performed with the presence of mesangial connective tissue inflammation as the dependent variable (no inflammation = 0; inflammation = 1). The independent variables included in the multivariate analysis were main pancreatic duct dilatation as reported by CT imaging, lesion length (≤2 cm or >2 cm), and lesion location (ampullary lesion, pancreatic head lesion, and duodenal lesion). The results of multivariate analysis showed that main pancreatic duct dilatation reported by CT imaging (OR=2.580, 95%CI:1.234-5.395, P=0.012) was a risk factor for inflammation of the pancreatic mesentery connective tissue, while lesion long diameter ≤2 cm (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 for inflammation of the pancreatic mesentery connective tissue. The results of multivariate analysis showed that 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) reported by CT imaging were independent risk factors for inflammation of the non-mesenteric connective tissue.

[0040] The segmentation methods for connective tissue ROI 1 in the LPD surgical area include: The CT venous phase images were preprocessed by adjusting the window width and window position and performing nonlinear filtering. On the pre-processed CT venous phase images, the TotalSegmentator segmentation model was used to segment the important anatomical structures during LPD surgery, and a cropping frame was determined that included the pancreas, SMV-PV axis, celiac trunk, common hepatic artery and its main branches. On the CT venous phase images within the cropping box, the abdominal visceral fat is locally segmented using the nnUNet segmentation framework; On the pre-processed CT venous phase images, global segmentation of abdominal visceral fat was performed using 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.

[0041] The nnUNet segmentation framework in the present invention performs segmentation of abdominal visceral tissue on a cropping frame containing the pancreas, SMV-PV axis, celiac trunk, common hepatic artery and its main branch structures. It is a local image segmentation based on CT images, which realizes refined segmentation focusing on key areas. However, compared with the segmentation of the global image, the local image may inevitably lose some image information, resulting in impaired segmentation accuracy. In order to compensate for the loss of segmentation accuracy caused by this partial information loss, the present invention also performs a segmentation of abdominal visceral tissue on the original CT image with the nnUNet segmentation framework. It is a global image segmentation based on the CT image. 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, thereby avoiding the loss of segmentation accuracy caused by information loss, and further improving the segmentation accuracy of the hierarchical modeling strategy, as follows: The method of spatially registering the local segmentation result of abdominal visceral fat with the global segmentation result of abdominal visceral fat and performing mask operation includes: Performing spatial registration on the global segmentation result of the abdominal visceral fat and the clipping frame to obtain the global segmentation result of the abdominal visceral fat in the clipping frame coordinate system; Extracting the global segmentation result of the abdominal visceral fat within the clipping frame from the global segmentation result of the abdominal visceral fat within the clipping frame coordinate system; Real-time performance of the nnUNet segmentation framework for local segmentation of abdominal visceral fat, and real-time performance of the nnUNet segmentation framework for global segmentation of abdominal visceral fat; Based on the segmentation accuracy of local segmentation and the segmentation accuracy of global segmentation, the global segmentation results of the abdominal visceral fat within the cropping frame and the local segmentation results of the abdominal visceral fat are subjected to mask operation in a weighted fusion manner to obtain the LPD surgical area connective tissue ROI 1; The expression of mask operation is: ; Where, It is the connective tissue of the LPD operation area. is the global segmentation result of the abdominal visceral fat within the cropping box, This is the local segmentation result of abdominal visceral fat. is the segmentation accuracy of the global segmentation, is the segmentation accuracy of local segmentation.

[0042] The present invention adopts a weighted fusion approach in mask operations, and introduces the nnUNet segmentation framework to evaluate the real-time segmentation accuracy of local and global segmentation. That is, the global segmentation model performance and local segmentation model performance of the nnUNet segmentation framework at the current moment are judged. The higher the performance, the higher the confidence of the segmentation result. Therefore, after normalization, the global segmentation result and the weight of the local segmentation result can be used to supplement the segmentation accuracy lost by the local segmentation through global segmentation, thereby achieving the purpose of "adding the global image segmentation result to the local image segmentation result, which is equivalent to adding the original image information to the local segmentation to reduce information loss, thereby avoiding the loss of segmentation accuracy caused by information loss, and further improving the segmentation accuracy of the hierarchical modeling strategy."

[0043] The gold standard segmentation of ROI 1 to ROI 6 was obtained by manual segmentation by radiologists with professional experience in abdominal imaging.

[0044] The real-time segmentation accuracy of the nnUNet segmentation framework for local and global segmentation in the present invention is measured using a training set obtained from collected case samples. That is, the error between the local and global segmentation results obtained by the nnUNet segmentation framework in the training set and the gold standard of ROI 1 segmentation obtained by manual segmentation by radiologists with professional experience in abdominal imaging is calculated. The higher the error, the lower the segmentation performance and the lower the confidence of the segmentation result.

[0045] The segmentation accuracy is quantified by the difference between the global segmentation result of the abdominal visceral fat within the cropping box and the local segmentation result of the abdominal visceral fat and the segmentation gold standard of ROI 1. The expression for quantification of segmentation accuracy is: ; ; Where, is the global segmentation result within the cropping box obtained by the nnUNet segmentation framework on the i-th image sample at the current moment, is the local segmentation result obtained by the nnUNet segmentation framework on the i-th image sample at the current moment, is the gold standard for segmentation of ROI 1 on the i-th image sample, is the mean square loss formula, and m is the total number of image samples, which is consistent with the number of cases in the training set.

[0046] The local segmentation results of the abdominal visceral fat were spatially registered with the global segmentation results of the abdominal visceral fat and masked to obtain the LPD surgical area connective tissue ROI 1 (recorded as Original when predicting the presence or absence of inflammation, as shown in Figure 14a to c).

[0047] Methods for extracting radiomic features of ROI 1 include: 107 radiomics features were extracted from the connective tissue ROI 1 of the LPD surgical area using the PyRadiomics component in the Python library; The SelectKBest function combined with chi-square test or variance analysis was used to screen out 20 important imaging features consisting of 1 shape feature, 4 first-order features and 15 texture features from 107 imaging features, such as Figure 8 As shown in the figure, it is used as the ROI 1 radiomics feature for discriminating the presence of inflammation in the pancreatic surgical area.

[0048] When predicting whether the connective tissue in the pancreatic mesentery of LPD patients has inflammation, 142 patients entered the training set, including 78 patients without inflammation and 64 patients with inflammation; 61 patients entered the validation set, including 34 patients without inflammation and 27 patients with inflammation. When predicting the degree of inflammation, 63 patients entered the training set, including 30 patients with mild inflammation and 33 patients with severe inflammation; 28 patients entered the validation set, including 14 patients with mild inflammation and 14 patients with severe inflammation. In addition, there were 16 patients in the external validation set composed of a prospective cohort, including 8 patients in the non-inflammation group, 4 patients in the mild inflammation group, and 4 patients in the severe inflammation group. The screening and grouping of patients are as follows. Figure 7 shown.

[0049] The method for constructing the inflammation presence discrimination model includes: Four radiomics models were constructed using logistic regression (LR), support vector machine (SVM), decision tree (DT), and random forest (RF), all with ROI1 radiomics features as input and inflammation presence as output. The performance of the four radiomics models was comprehensively evaluated using sensitivity (SEN), specificity (SPE), accuracy (ACC), precision (PRE), F1 score, and area under the curve (AUC) of the receiver operating characteristic curve (ROC). When predicting the presence or absence of inflammation, based on the 20 important radiomic features extracted from ROI1, four machine learning algorithms, LR, SVM, DT, and RF, were used to construct a radiomics model (Rad-score1), namely the radiomics model. Figure 9 As shown in Figure 2, 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 validation set are shown. Figure 10 As shown, Figure 10 The radar chart is drawn based on the above data. The closer the performance parameter point in the radar chart is to the outermost circle, the better the model performs in this parameter; the larger the area covered by the radar chart, the stronger the overall performance of the model. In the training set, Figure 9 and Figure 10 The results showed that among the four radiomics models, the RF model outperformed the other three models in all other performance parameters, with an AUC of 0.982 (95% CI: 0.963-0.994) and an accuracy of 0.937, except for a slightly lower sensitivity than the other three machine learning algorithms. In the validation set, the RF model outperformed the other three models in all other performance parameters, with an AUC of 0.807 (95% CI: 0.691-0.913) and an accuracy of 0.770, except for a lower sensitivity than the other three machine learning algorithms. The DT model performed worse than the RF model, with an AUC of 0.752 (95% CI: 0.637-0.856) and an accuracy of 0.689. Among the radiomics models constructed using the four machine learning algorithms, LR and SVM performed poorly. Therefore, the RF algorithm was ultimately selected to construct the radiomics model for predicting the presence or absence of inflammation (i.e., the model with the best overall performance among the four radiomics models).

[0050] The model with the best comprehensive performance was selected from the four radiomics models. The ROI 1 radiomics features and clinical risk factors were used as combined inputs, and the inflammation presence results were used as output to construct an inflammation presence discrimination model.

[0051] When predicting the presence or absence of inflammation, the performance comparison of the clinical model (RF model trained with clinical risk factors as input and inflammation presence results as output), the best radiomics model (i.e., the model with the best overall performance among the four radiomics models), and the fusion model (i.e., the inflammation presence discrimination model) is shown in Figure 11 and Figure 12 , where clinical is the clinical model, rad is the optimal radiomics model, and all is the fusion model. In the validation and external validation sets, the fusion model performed best, with AUCs of 0.826 and 0.734 and accuracies of 0.770 and 0.688, respectively. The radiomics model performed second best, with AUCs of 0.807 and 0.766 and accuracies of 0.770 and 0.625, respectively. However, there was no statistically significant difference in AUC between the fusion and radiomics models in the validation and external validation sets (DeLong test, P values ​​of 0.360 and 0.400, respectively). The clinical model had the poorest ability to predict pancreatic mesangial connective tissue inflammation in the training, validation, and external validation sets, with AUCs of 0.634, 0.678, and 0.539, respectively. The difference in AUC between the clinical and fusion models was statistically significant (DeLong test, P values ​​less than 0.001, respectively). In the validation set, 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; the calibration curve of the fusion model showed good consistency between the predicted results and the actual situation with or without inflammation, such as Figure 12 shown.

[0052] Methods that use ROI 1 radiomics features and clinical risk factors as combined input include: The attention weights of ROI 1 radiomic features and clinical risk factors were obtained by cross-attention allocation, where: Attention weights of ROI 1 radiomics features , where is the attention weight vector of ROI 1 radiomics features, is the query vector derived from clinical risk factors, The key vector from the radiomics features of ROI 1, is the value vector from the radiomics feature of ROI 1, for The vector dimension of Attention weighting of clinical risk factors , where is the attention weight vector of clinical risk factors, is the query vector from the radiomics features of ROI 1, Key vectors derived from clinical risk factors, is the value vector from clinical risk factors, for The vector dimension of The attention weights of ROI 1 radiomics features and clinical risk factors were combined with the ROI 1 radiomics features and clinical risk factors to obtain the ROI 1 radiomics enhanced features. and clinical risk factor enhancement features ; In order to improve the computational efficiency of the inflammation presence discrimination model, the present invention quickly captures key information in the combined input obtained by combining ROI 1 radiomics features and clinical risk factors, and adopts a cross-attention mechanism. This can capture the correlation between the two modal features between ROI 1 radiomics features and clinical risk factors, and realize information interaction between the two modal features. At the same time, in the process of realizing information interaction between the two modal features, attention weights are assigned to the two modal features according to the importance of the interactive information, so as to give high weights to features in ROI 1 radiomics features that have high correlation with clinical risk factors, and give high weights to features in clinical risk factors that have high correlation with ROI 1 radiomics features, thereby highlighting the key information related to clinical risk factors in ROI 1 radiomics features, and highlighting the key information related to ROI 1 radiomics features in clinical risk factors. This achieves the interaction of multimodal features, effectively establishes complementary connections between different modalities, helps the model identify and utilize the correlation between different modalities, and improves the efficiency and performance of the inflammation presence discrimination model.

[0053] Evaluate the differences between ROI 1 radiomics enhancement features and clinical risk factor enhancement features and ROI 1 radiomics features and clinical risk factors, and modify ROI 1 radiomics enhancement features based on the differences and clinical risk factor enhancement features Generalization, obtaining comprehensive radiomics features of ROI 1 and comprehensive characteristics of clinical risk factors ; Furthermore, in order to avoid the multimodal interaction between ROI 1 radiomics features and clinical risk factors due to the cross-attention mechanism, the present invention causes the key information related to clinical risk factors in ROI 1 radiomics features to be too prominent, and the key information related to ROI 1 radiomics features in clinical risk factors to be too prominent, so that the model can only focus on the interactive information, while the remaining information in clinical risk factors and ROI 1 radiomics features except the interactive information will be ignored due to the small assigned weight. The remaining information also contains information that is beneficial to the discrimination of the presence of inflammation, which will cause the loss of feature information in the discrimination process.

[0054] Therefore, the present invention limits the degree of multimodal interaction, that is, on the basis of the ROI 1 radiomics features obtained by highlighting the key information related to the clinical risk factors through cross-attention weighting (i.e., ROI 1 radiomics enhanced features), the original ROI 1 radiomics features are reintroduced, thereby supplementing the feature information lost after the ROI 1 radiomics features are enhanced through cross-attention, and obtaining the ROI 1 radiomics comprehensive features; and on the basis of the clinical risk factors obtained by highlighting the key information related to the ROI 1 radiomics features through cross-attention weighting (i.e., clinical risk factor enhanced features), the original clinical risk factors are reintroduced, thereby supplementing the feature information lost after the clinical risk factors are enhanced through cross-attention, and obtaining the clinical risk factor comprehensive features.

[0055] Specifically, in the process of supplementing the feature information lost in ROI 1 radiomics features due to cross-attention enhancement, the strength of the supplement is related to the similarity between the enhanced features of ROI 1 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, a higher supplementary strength should be given to the original ROI 1 radiomics feature. The present invention gives a higher weight to the original ROI 1 radiomics feature to correspond to the supplementary strength, that is, when the similarity between the ROI 1 radiomics enhancement feature and the original ROI 1 radiomics feature is greater than that between the original ROI 1 radiomics feature and the original ROI 1 radiomics feature, the higher the supplementary strength is. The lower the value, the higher the weight is given to the original ROI 1 radiomics features to supplement the large amount of lost feature information. When the similarity between the ROI 1 radiomics enhanced features and the original ROI 1 radiomics features is The higher the value, the lower the weight is given to the original ROI 1 radiomics features to supplement the small amount of lost feature information.

[0056] Similarly, in the process of supplementing the characteristic information lost after the clinical risk factors are enhanced due to cross-attention, the strength of the supplement is related to the similarity between the enhanced characteristics of the clinical risk factors and the original clinical risk factor characteristics. Correlation, similarity between enhanced features of clinical risk factors and original clinical risk factor 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, a higher supplementary strength should be given to the clinical risk factors. The present invention gives a higher weight to the original clinical risk factors to correspond to the supplementary strength, that is, when the similarity between the enhanced features of the clinical risk factors and the original clinical risk factors is greater than that of the original clinical risk factors, the higher the supplementary strength should be. The lower the value, the higher the weight is given to the original clinical risk factors to supplement the large amount of lost feature information. When the clinical risk factor enhancement feature is similar to the original clinical risk factor, The higher it is, the lower the weight is given to the original clinical risk factors to supplement the small amount of missing feature information.

[0057] The radiomics comprehensive features of ROI 1 and comprehensive characteristics of clinical risk factors The combination constitutes the combined input; The combined input is: ; Where, For combined input, is the comprehensive radiomics feature of ROI 1, is a comprehensive feature of clinical risk factors. is the channel splicing function, is the radiomics feature of ROI 1, Clinical risk factor characteristics, For ROI 1 radiomics enhancement features, Enhanced characterization of clinical risk factors, for The attention weight, for The attention weight, is the similarity function.

[0058] Clinical risk factors include lesion location, lesion size, and the presence or absence of main pancreatic duct dilatation.

[0059] The results for the presence of inflammation included inflammation and no inflammation.

[0060] like Figure 13 As shown in the figure, the SHAP (Shapley Additive exPlanations) framework was used to perform interpretability analysis on the inflammation presence discrimination model.

[0061] The extracted radiomics features are limited to first-order features, shape features, and texture features, which are well-correlated with inflammatory pathological mechanisms. First-order features reflect the grayscale distribution of individual voxels (or pixels) and can capture the homogeneity of tissue density. These include the 90th percentile, mean, and range. Shape features reflect the appearance and size of the ROI and can quantify anatomical distortions, such as minor axis length, shape elongation, and maximum 3D diameter. Texture features can analyze tissue heterogeneity and include the following five categories: Gray Level Co-occurrence Matrix (GLCM), Gray Level Dependence Matrix (GLDM), Gray Level Run Length Matrix (GLRLLM), Gray Level Size Zone Matrix (GLSZM), and Neighborhood Gray Tone Difference Matrix (NGTDM). In the model predicting the presence or absence of inflammation, the Shapley Additive Explanations (SHAP) of the top 20 radiomics features that contribute most to the model are as follows: Figure 13 When the SHAP value of the radiomics feature is positive, the model prediction probability increases, while a negative value decreases the model prediction probability.

[0062] The discriminant model for the presence of inflammation was constructed using Random Forest (RF).

[0063] In the fusion model constructed by combining risk factors and imaging omics features, the present invention uses the TotalSegmentator segmentation model and the nnUNet segmentation framework to segment the connective tissue ROI 1 of the LPD surgical area and extract the imaging omics features of ROI 1. Retrospective studies are used to discover clinical risk factors, and screening operations are supported to ensure the accuracy of the fusion model in distinguishing the presence of inflammation. At the same time, clinical risk factors and imaging omics features are spliced ​​using the attention mechanism to highlight key features, allowing the fusion model to quickly capture key features, thereby ensuring the efficiency of the fusion model in distinguishing the presence of inflammation.

[0064] The above embodiments are merely exemplary embodiments of the present application and are not intended to limit the scope of the present application. The scope of protection of the present application is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present application within the essence and scope of protection of the present application, and such modifications or equivalent substitutions shall also be deemed to fall within the scope of protection of the present application.

Claims

1. A method for identifying the presence of inflammation in the pancreatic surgical area by integrating fat-related imaging features, characterized in that: The following steps are involved: Clinical risk factors for the presence of pancreatic surgical site inflammation were determined by statistical methods; On CT venous phase images, the TotalSegmentator segmentation model and the nnUNet segmentation framework were combined to segment the LPD surgical area connective tissue ROI 1. Radiomic features of ROI 1 for discriminating the presence of pancreatic inflammation were extracted from the LPD surgical area connective tissue ROI 1. Through multiple machine learning algorithms, an inflammation presence discrimination model for discriminating the presence of inflammation in the pancreatic surgical area was constructed based on the ROI 1 imaging features and clinical risk factors.

2. The method for determining the presence of pancreatic surgical area inflammation by integrating fat-related imaging features according to claim 1, characterized in that: The method for determining the clinical risk factors includes: Basic clinical data, laboratory test data, preoperative abdominal enhanced CT data, pathological data, and perioperative outcome data were obtained; The Shapiro-Wilk test was used to determine whether the measurement data in basic clinical data, laboratory test data, preoperative abdominal enhanced CT data, pathological data, and perioperative outcome data conformed to normal distribution. The continuous variables in the measurement data that conformed to the normal distribution were described by means of mean and standard deviation. The two independent sample t-test was used for comparison between the two groups when the variances were equal; otherwise, the Mann-Whitney U test was used. The variables that did not conform to the normal distribution in the measurement data were described using the median, 25th percentile, and 75th percentile, and the Mann-Whitney U test was used for comparison between groups; The enumeration data of basic clinical data, laboratory examination data, preoperative abdominal enhanced CT data, pathological data, and perioperative outcome data were described as number of cases and percentages. The chi-square test was used for comparison between groups. For comparisons between the two groups, Fisher's exact test was used when theoretical frequencies existed, and the continuity-corrected chi-square test was used when 1 < theoretical frequency < 5 existed; When multiple groups were compared without a correction formula, Fisher's exact test was used when the theoretical frequency was <1 or when the number of cells with 1 ≤ theoretical frequency <5 exceeded 20% of the total number of cells; Stepwise regression was used to screen the variables with statistical significance in univariate analysis, and multivariate binary logistic regression analysis was used to identify clinical risk factors.

3. The method for determining the presence of pancreatic surgical area inflammation by integrating fat-related imaging features according to claim 1, characterized in that: The segmentation method of the LPD surgical area connective tissue ROI 1 includes: The CT venous phase images were preprocessed by adjusting the window width and window position and performing nonlinear filtering. On the pre-processed CT venous phase images, the TotalSegmentator segmentation model was used to segment the important anatomical structures during LPD surgery, and a cropping frame was determined that included the pancreas, SMV-PV axis, celiac trunk, common hepatic artery and its main branches. On the CT venous phase images within the cropping box, the abdominal visceral fat is locally segmented using the nnUNet segmentation framework; On the pre-processed CT venous phase images, global segmentation of abdominal visceral fat was performed using the nnUnet segmentation framework; The local segmentation result of the abdominal visceral fat is spatially registered and masked with the global segmentation result of the abdominal visceral fat to obtain the LPD operation area connective tissue ROI 1.

4. The method for determining the presence of pancreatic surgical area inflammation by integrating fat-related imaging features according to claim 3, characterized in that: The method for extracting radiomics features of ROI 1 includes: 107 radiomics features were extracted from the connective tissue ROI 1 of the LPD surgical area using the PyRadiomics component in the Python library; Twenty important imaging features consisting of one shape feature, four first-order features, and 15 texture features were screened out from 107 imaging features using the SelectKBest function combined with the chi-square test or analysis of variance. These features were used as the ROI 1 imaging features for discriminating the presence of inflammation in the pancreatic surgical area.

5. The method for determining the presence of pancreatic surgical area inflammation by integrating fat-related imaging features according to claim 4, characterized in that: The method for constructing the inflammation presence discrimination model includes: Four radiomics models were constructed using logistic regression, support vector machine, decision tree, and random forest, all with ROI 1 radiomics features as input and inflammation presence results as output; The performance of the four radiomics models was comprehensively evaluated using sensitivity, specificity, accuracy, precision, F1 score, and area under the receiver operating curve. The model with the best comprehensive performance was selected from the four radiomics models. The ROI 1 radiomics features and clinical risk factors were used as combined inputs, and the inflammation presence results were used as output to construct the inflammation presence discrimination model.

6. The method for determining the presence of pancreatic surgical area inflammation by integrating fat-related imaging features according to claim 5, characterized in that: The method using ROI 1 radiomics features and clinical risk factors as combined input includes: The attention weights of ROI 1 radiomic features and clinical risk factors were obtained by cross-attention allocation, where: Attention weights of the radiomics features of ROI 1 , where is the attention weight vector of ROI 1 radiomics features, is the query vector derived from clinical risk factors, The key vector from the radiomics features of ROI 1, is the value vector from the radiomics feature of ROI 1, for The vector dimension of Attention weighting of the clinical risk factors , where is the attention weight vector of clinical risk factors, is the query vector from the radiomics features of ROI 1, Key vectors derived from clinical risk factors, is the value vector from clinical risk factors, for The vector dimension of The attention weights of ROI 1 radiomics features and clinical risk factors were combined with the ROI 1 radiomics features and clinical risk factors to obtain the ROI 1 radiomics enhanced features. and clinical risk factor enhancement features ; Assess the differences between ROI 1 radiomics enhancement features and clinical risk factor enhancement features and ROI 1 radiomics features and clinical risk factors, and modify ROI 1 radiomics enhancement features based on the differences and clinical risk factor enhancement features The generalization of ROI 1 radiomics comprehensive features was obtained. and comprehensive characteristics of clinical risk factors ; The radiomics comprehensive features of ROI 1 and comprehensive characteristics of clinical risk factors Combining to form said combined input; The combined input is: ; Where, For combined input, is the comprehensive radiomics feature of ROI 1, is a comprehensive feature of clinical risk factors. is the channel splicing function, is the radiomics feature of ROI 1, Clinical risk factor characteristics, For ROI 1 radiomics enhancement features, Enhanced characterization of clinical risk factors, for The attention weight, for The attention weight, is the similarity function.

7. The method for determining the presence of pancreatic surgical area inflammation by integrating fat-related imaging features according to claim 2, characterized in that: The clinical risk factors include lesion location, lesion size, and the presence or absence of main pancreatic duct dilatation.

8. The method for determining the presence of pancreatic surgical area inflammation by integrating fat-related imaging features according to claim 5, characterized in that: The results of the presence of inflammation include inflammation and no inflammation.

9. The method for determining the presence of pancreatic surgical area inflammation by integrating fat-related imaging features according to claim 1, characterized in that: The SHAP framework was used to perform interpretability analysis on the inflammation presence discriminant model.

10. The method for determining the presence of pancreatic surgical area inflammation by integrating fat-related imaging features according to claim 5, characterized in that: The inflammation presence discriminant model is constructed by random forest.

Images