Radiomics analysis based on CT’s greater omental caking for predicting pathological grading of pseudomyxoma peritonei

The objective of this study was to predict the preoperative pathological grading and survival period of Pseudomyxoma peritonei (PMP) by establishing models, including a radiomics model with greater omental caking as the imaging observation index, a clinical model including clinical indexes, and a combined model of these two. A total of 88 PMP patients were selected. Clinical data of patients, including age, sex, preoperative serum tumor markers [CEA, CA125, and CA199], survival time, and preoperative computed tomography (CT) images were analyzed. Three models (clinical model, radiomics model and combined model) were used to predict PMP pathological grading. The models’ diagnostic efficiency was compared and analyzed by building the receiver operating characteristic (ROC) curve. Simultaneously, the impact of PMP’s different pathological grades was evaluated. The results showed that the radiomics model based on the CT’s greater omental caking, an area under the ROC curve ([AUC] = 0.878), and the combined model (AUC = 0.899) had diagnostic power for determining PMP pathological grading. The imaging radiomics model based on CT greater omental caking can be used to predict PMP pathological grading, which is important in the treatment selection method and prognosis assessment.

Image and data. All patients had a plain abdominal and pelvic CT scan before the operation using 64-slice spiral CT (Siemens Healthcare, Germany) with a thickness of 5 mm, an interval of 5 mm, and a pitch of 1 to 1.2. All patients were prepared for gastrointestinal tract before examination: fasting and water deprivation for 8 h. CT images were stored in DICOM format. The images were imported into ITK-SNAP software (version 3.8.0), and the tumor boundary as manually produced to determine the tumor region of interest (ROI). The thickening of the omentum which is recognized by CT was defined as omental caking. The standard of tracing greater www.nature.com/scientificreports/ omental caking was used to select the image that could display the largest greater omental caking; the ROI area enveloped the greater omental caking, and the unit was square centimeter. The ROI area of greater omental caking is shown in Fig. 2. All cases were retrospectively studied by two radiologists, working independently. Meanwhile, the accuracy of lesions and relevant data were inspected by a specialist.
Radiomics feature extraction. The Dr. Wise Multimodal Research Platform (https:/keyan.deepwise. com; Beijing Deepwise & League of PHD Technology Co., Ltd, 193 Beijing, China) was used for the CT's feature extraction. Approximately 1218 features were extracted from the CT lesion's ROI extracted and categorized into seven categories: first order features, shape based features, gray-scale co-occurrence Matrix (GLCM) features, gray-level size zone matrix (GLSZM) features, gray-level run length matrix (GLRLM) features, gray-level distance-zone matrix (GLDM), and neighboring gray level dependence matrix 16 .
The results of training sets and validation sets were tested by the fivefold cross validation method. To study the correlation between clinical factors and greater omental caking, logistic regression models were deployed. To build the radiomics model, and six kinds of feature-screening methods were used: those that included the F-Test, Pearson Correlation Coefficient, Mutual Information, L1-based models, tree-based models, and Recursive Feature Elimination. We only used one of these methods each time in model construction. Finally, a method based linear models penalized with the L1 norm was selected for feature screening and it had the best diagnostic performance. RAD-SCORE was generated by analyzing logistic regression weighted by their coefficients of the features. We used multivariate logistic regression to construct the clinical-radiomics combined model; meanwhile, it was presented in the nomogram chart 16 . Statistical analysis. Univariate logistic regression and multivariate logistic regression were used to establish the clinical model. Then, the backward stepwise selection method was applied. In evaluation of each model's performance, we utilized the area under the curve (AUC) with 95% confidence intervals (95% CIs). The DeLong test was used to determine whether all the models' AUC values were significant. Finally, the nomogram figure of the clinical-radiomics model was established. To confirm the clinical usefulness of the three models, decision curve analysis was conducted. A Cox regression model with the Kaplan-Meier method was used for survival analysis. R software (version 4.0.2) and SPSS (version 23.0) were applied for analysis.

Generated statement.
No animal studies are presented in this manuscript.
Studies involving human subjects. Generated Statement: The studies involving human participants were reviewed and approved by the Ethics Committee of Aerospace Center Hospital. The patients provided written informed consent to participate in this study.
Inclusion of identifiable human data. Generated Statement: No potentially identifiable human images or data are presented in this study.

Results
Patient characteristics. Of the 88 patients, 60 (68%) presented low-grade PMP, and 28 (32%) were diagnosed with high-grade PMP. The primary sites of 88 patients enrolled in our study are the appendix. Table 1 shows all patient features.
Clinical model. Among the clinical indicators, including sex, age, and preoperative serum tumor markers (CEA, CA199, and CA125), CA125 was an important factor that prompted the pathological grading of PMP  Clinical-radiomics combined model. The combined clinical-radiomics model showed the best diagnostic performance, with an AUC of 0.92 in the training test and 0.90 in the validation test, respectively. The AUC, accuracy, sensitivity, and specificity of the three models are listed in Table 2. The ROC curves are presented in Fig. 3. The nomogram of the clinical-radiomics combined model is displayed in Fig. 4. From the ROC curves, the cut-off values were got in Fig. 3. The cut-off value of radiomics model is − 0.414 and the corresponding nomogram value is 25 points. The cut-off value of clinical model is 92.5 and the corresponding nomogram value is 10 points. So that, the cut-off value of nomogram is 35 points.
Decision curve. The decision curves of the three models are presented in Fig. 5. From the decision curve, the clinical-radiomics combined model was more beneficial than other models in predicting pathological grading.
Survival time. In this study, the follow-up time was 0-2000 days; the median follow-up time for high-grade PMP was 615 days, and the median follow-up time for low-grade PMP was 750 days. There was a significant difference in the follow-up time between the high-grade and low-grade PMP groups (P = 0.039). The survival rate curve is shown in Fig. 6.

Discussion
PMP is a rare clinical entity that is characterized by a significant amount of mucus-jelly substance that widely spreads and settles in the abdominal cavity. The onset of the disease is insidious; the clinical symptoms are atypical, and early diagnosis is difficult. For the differential diagnosis of PMP pathological grading, the sensitivity of physical examination and imaging examination is not high. Previous studies have indicated that CT identification accuracy and quantification of various peritoneal surface malignancies vary in different primary tumors of origin 17,18 . Presently, CT is also commonly used as a     www.nature.com/scientificreports/ Our research center collected a total of 124 PMP patients from 2014 to 2019. By analyzing the characteristics of the patients' CT images, we found that 88% of PMP patients had greater omental caking. Studies have shown that the omentum is composed of two layers of peritoneum, which contains blood vessels, lymphatic vessels, lymph nodes, and fatty tissue. The omentum is not only a barrier to disease spread but also a pathway. Hence, the omentum is vulnerable to disease invasion 23 . Tumor cells with PMP exocrine function are widely planted in the peritoneum, causing a significant amount of mucus to accumulate in the abdominal cavity. The mucus protein content is high, and the mucin/cell ratio can be as high as 1:1000, which is relatively viscous 24 . Long-term stimulation of the greater omentum can easily cause fibrosis and the formation of greater omental caking. Greater omental caking is a common imaging feature of PMP, which is one of the reasons why we chose it as a radiomics analysis. Additionally, previous studies have shown that CT is more sensitive and accurate for detecting 0 zone lesions of PMP, which are entire transverse colon and greater omentum lesions 25 . Therefore, we selected greater omental caking as a target of radiomics analysis.
Some studies have shown that the pathological manifestation of jelly like mucus in the abdominal cavity of PMP is simple acellular mucus. Its density, which is the CT value, is higher than that of conventional ascites and other cystic lesions but lower than the soft tissue density. The cell-free mucus material showed a low-density area in a plain CT scan 20 . Different grades of PMP have similar clinical symptoms and similar imaging manifestations on plain CT scans. Traditional imaging can only provide general information, such as tumor size, shape, or density, and has limited value for distinguishing the pathological grading of PMP. Radiomics transforms the quantitative features in high-through put extracted images into data information and establishes predictive models, which can microscopically reflect tumor heterogeneity and are closely related to tumor histological components and the degree of differentiation. Hence, radiomics provides is effective in the differential diagnosis 26 . Radiomics provided a new opportunity to analyze whether the mucus content in PMP greater omental caking was related to its pathological grade and whether the mucus content could be reflected by CT radiomics analysis to predict the pathological grading.
The study results showed that the characteristics of the greater omental caking CT images were significantly different between high-grade and low-grade PMP. Among them, wavelet-HHL_gldm_Small Dependence Low Gray Level Emphasis (SDLGLE), which measures the small dependence's joint distribution with lower gray-level values, was higher in low grade, indicating that the low-density areas in the low-grade PMP's greater omental caking were more dispersed and corresponded to histopathology; it stated that mucus components were more widely distributed, and cells and fiber components were less in intensity. Additionally, the parameter wavelet-HHH_glszm_Gray Level Non-Uniformity (GLN), which measures the variability of the gray-level intensity values in the image, had a lower value in the low-grade PMP, indicating poor uniformity of the low-grade PMP's greater omental caking in histopathology, and there were more mucus components inside the lesion and fewer cells and fibers. Wavelet-HLL_glszm_SmallAreaLowGrayLevelEmphasis (SALGLE), which measures the proportion in the image of the joint distribution of smaller size zones with lower gray-level values, was higher in low-grade PMP, suggesting that there were more low-density areas in the low-grade PMP's greater omental caking, and its pathology showed that there were more mucus components and fewer cells and fibers.
The radiomics model of this study showed that there were more low-density areas of the low-grade PMP greater omental caking; there were more mucus components, and there were fewer cell and fiber components. The greater omental caking radiomics features showed significant differences between high-grade and low-grade PMP. The analysis showed that this was related to the pathological and cytological differences between those two. In low-grade PMP, the secretory function of tumor epithelial cells was strong, and a large number of mucus vacuoles could be seen in columnar cells, while in high-grade PMP, the cells were severely dysplastic, which severely damaged the columnar secretory function of the cell, thus reduced cellular mucus 27 .
In addition, this study found that serum CA125 levels were correlated with different pathological grade. In this study, the patient's sex, age, preoperative serum tumor markers (CEA, CA199, and CA125), and other relevant clinical data were collected to establish a clinical model. It was found that serum CA125 was meaningful in PMP pathological grading. The clinical model had a sensitivity of 0.43 and a specificity of 0.73, indicating that the serum CA125 level had value in predicting PMP's different pathological grading. www.nature.com/scientificreports/ In 1981, Bast obtained the monoclonal antibody OC25 from papillary serous cystic ovarian cancer. The tumor-associated antigen recognized by this antibody was CA125 28 . In 2001, the gene CA125 was cloned and named MUC16 29 ; CA125 is a large transmembrane glycoprotein that exists in normal peritoneal mesothelial cells and Mullerian epithelial cells 30,31 . According to the literature reports, serum CA125 levels in humans are simultaneously affected by CA125 produced by primary tumor cells and peritoneal mesothelial cells. Because of cell connections and the basement membrane, it cannot enter the blood circulation, so normal people's serum CA125 levels are low. PMP's typical clinical manifestations are mucinous ascites and extensive peritoneal implantation 2,32 . Under influence of malignant tumors, the normal cell membrane barrier is destroyed, leading to the release of CA125 into the blood. The upregulation of the MUC16's expressionalso causes increase of the serum CA125 level 33 . Additionally, recent studies have shown that CA125 participates in the process of peritoneal implantation by mediating the adhesion of free tumor cells in the abdominal cavity and mesothelin on the surface of peritoneal mesothelial cells 34,35 . High-grade PMP is rich in peritoneal mesothelial cells, so its serum CA125 level is higher, while low-grade PMP is the opposite. In Ross's study, CA125 was the only marker to predict complete cytoreduction independently and to be associated with long-term survival. Patients with elevated CA125 had a lower 3-year survival rate 7 . However, CA125 is also affected by peritonitis, ovarian swelling and so on 29 . Therefore, preoperative serum CA125 levels and different pathological grade have some correlation.
CEA is a broad-spectrum tumor marker, and the detection of its level through continuous follow-up can reflect the curative effect and prognosis of malignant tumors after surgery. CA199 is a glycoprotein tumor antigen. It is a gastrointestinal tumor-related antigen. In our study, we also tried to identify the pathological grading of PMP by CEA and CA199. These tumor markers were also beneficial in the prognosis assessment, and a normal level indicated an improved prognosis 36 . SHIGEKI 37 analyzed 156 PMP cases and concluded that optimal cutoffs were 18 ng/ml for CEA and 89 U/ml for CA199. The AUCs-ROC were 0.68 and 0.69 for CEA and CA199. However, the Ross study concluded that there was no statistically significant association between elevated CEA, CA199, and tumor histology 7 . Theses results were consistent with the results of our study. The significance of CEA and CA199 are controversial in predicting pathological grading of PMP before surgery.
In our study, the combined clinical-radiomics model showed the best diagnostic efficacy by comparing the ROC curves. According to the decision curve, the clinical-radiomics combined model was more beneficial in a wide range of high risk threshold than the clinical and radiomics models alone in predicting high grade PMP. The cut-off value of nomogram is 35 points. That is, nomogram higher than 35 points indicates high-grade PMP, and nomogram lower than 35 points indicates low-grade PMP. Improvement of accuracy in models combining radiomics with clinical factors has previously been reported in many studies 16,38 .
The survival curve shows that the survival rate of low-grade PMP is higher than that of high-grade PMP. This result is related to PMP's pathological basis. The internationally recognized PMP pathological grading is mainly based on the number of tumor cells and the shape of cancer nests, cell atypia, mitotic figures, and surrounding forms of invasion 32 . The pathological feature of low-grade PMP is mucus. There are many pools and few tumor epithelial cells in the mucus. The tumor cells are columnar or flat and arranged in a single layer or in the form of cords, islands, and clusters. Tumor cells are mildly atypical, and the nuclei are small and regular. There is push-like infiltration of other organs. The pathological features of high-grade PMP are abundant tumor epithelial cells, which can be arranged in strips, islands or cribriform; tumor cells are severely atypical (at least focally atypical), with large nuclei, obvious nucleoli, and vesicular chromatin. Mitotic figures are more common and can destructively infiltrate the internal organs of the abdominal cavity and significantly promote the fibro interstitial response 39 . Low-grade tumors have optimal cytological characteristics, while high-grade tumors have many cells and high heterogeneity. Choudry 40 retrospectively analyzed the clinicopathological data of 310 PMP patients who underwent CRS-HIPEC and divided the patients into 4 groups according to the tumor epithelial cell density: acellular mucus, low-cell density (< 2%), medium-cell density (2% to 19%), and high-cell density (> 20%). The results showed that compared to patients with acellular mucus, PMP patients with medium-cell or high-cell density had a shorter progression-free survival period. Horvath 41 also obtained similar results. The above two studies suggested that with the increase in tumor cell density, PMP patients had stronger tumor biological activity and a higher risk of disease progression. The biological behavior of low-grade tumor pushing/ expanding infiltration was more benign than invasive/destructive infiltration 32 . The biological behavior of highgrade tumors is often more malignant; the tumor tissue quickly proliferates, and it easily extensively involves the surrounding organs, making it difficult to easily remove and remetastasize 42 . Therefore, the survival probability of high-grade PMP is lower than that of low-grade PMP.
This study has the following advances: (1) Previous CT imaging studies were limited to PMP morphological studies. This study is the first report on PMP pathological grading through CT imaging. (2) This study used CT radiomics characteristics and a radiomics model, which were different from the indexes of most studies. Our study proved that combining RAD-SCORE and a clinical model can improve the diagnostic accuracy and provide clinical benefit of CT for PMP pathological grading. (3) The nomogram of the clinical-radiomics combined model shows several disease risk elements, and it presents a risk score for predicting prognosis, which is more convenient and distinct.
There were a few limitations for this study. (1) This study only included the characteristics of greater omental caking on plain CT scanning with the thickness of 5 mm. In the future, it will be useful to add thin section scanning and enhanced CT features for multimodal studies. (2) This study was performed with a small sample size at a single institution. In the future, it will be valuable to obtain more evidence by multicenter verification with a larger sample size. Due to the sample size, a small survival difference between patients with abnormal and normal CEA and CA199 levels was not detected. As the data continue to accumulate, further analysis needs to be implemented. (3)

Summary
In this study, a radiomics model based on CT greater omental caking was established to predict PMP pathological grading, and the results proved that this model has high diagnostic efficiency. Additionally, the nomogram can be used to predict PMP pathological grading by integrating image characteristics and clinical indicators. The chart is simple, intuitive, and easy to grasp, which can enhance imaging physicians' diagnostic confidence; clinicians can receive valuable guidance with a certain degree of preoperative prediction of PMP at the pathological level and specificity of treatment methods.

Data availability
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.