Abstract
Peritoneal metastasis (PM) is a frequent manifestation of advanced abdominal malignancies. Accurately assessing the extent of PM before surgery is essential for patients to receive optimal treatment. Therefore, we propose to construct a deep learning (DL) model based on enhanced computed tomography (CT) images to stage PM preoperatively in patients. All 168 patients with PM underwent contrast-enhanced abdominal CT before either open surgery or laparoscopic exploration, and peritoneal cancer index (PCI) was used to evaluate patients during the surgical procedure. DL features were extracted from portal venous-phase abdominal CT scans and subjected to feature selection using the Spearman correlation coefficient and LASSO. The performance of models for preoperative staging was assessed in the validation cohort and compared against models based on clinical and radiomics (Rad) signature. The DenseNet121-SVM model demonstrated strong patient discrimination in both the training and validation cohorts, achieving AUC was 0.996 in training and 0.951 validation cohort, which were both higher than those of the Clinic model and Rad model. Decision curve analysis (DCA) showed that patients could potentially benefit more from treatment using the DL-SVM model, and calibration curves demonstrated good agreement with actual outcomes. The DL model based on portal venous-phase abdominal CT accurately predicts the extent of PM in patients before surgery, which can help maximize the benefits of treatment and optimize the patient’s treatment plan.
Graphical abstract
![](http://media.springernature.com/lw685/springer-static/image/art%3A10.1007%2Fs10585-023-10235-5/MediaObjects/10585_2023_10235_Figa_HTML.png)
Similar content being viewed by others
Avoid common mistakes on your manuscript.
Introduction
Peritoneal metastasis (PM) is the most common metastasis in advanced abdominal tumors, with a median survival of no more than 12 months in some patients, and a worse prognosis than other metastasis, such as the liver or lung [1, 2]. In recent years, cytoreductive surgery (CRS) combined with hyperthermic intraperitoneal chemotherapy (HIPEC) has been widely accepted for treatment of PM, emerging as the most effective approach to prolonging survival in the majority of cases [3, 4]. CRS + HIPEC has been endorsed by the Peritoneal Surface Oncology Group International (PSOGI) as the established standard of care for patients diagnosed with peritoneal pseudomyxoma, malignant peritoneal mesothelioma, and colorectal peritoneal metastases. Furthermore, it is the recommended therapeutic approach for patients with peritoneal metastases associated with gastric and ovarian cancers [5,6,7,8,9,10].
However, not all patients may experience positive outcomes from CRS + HIPEC treatment. The peritoneal carcinomatosis index (PCI), together with the cytoreductive completeness score (CCS), has been shown to be a major determinant of overall patient survival [11,12,13]. Patients with severe PM often have a relatively poor prognosis and are unlikely to achieve a clinical cure after CRS treatment (CCS of 0 or 1). For instance, in colorectal cancer patients, PCI > 20 is considered a contraindication to CRS + HIPEC treatment [14], while gastric cancer patients with normal tumor markers and PCI ≤ 16 are more likely to benefit from surgical treatment [15]. Currently, the gold standard for assessing PCI is laparoscopic [16], and some patients with extensive metastasis may have to forgo further surgical treatment after the invasive examination. Open-close surgery has been reported in up to 23.4% of cases [17].
Imaging examinations such as CT and MRI have become common preoperative procedures for patients with PM. While they can provide information about the presence and extent of metastasis, their low sensitivity may result in discrepancies between the imaging results and the actual condition of the patient. As a result, CT-PCI and MRI-PCI demonstrate limited ability to stratify patients for selection when compared to surgical PCI (S-PCI), with a concordance index of only 0.47–0.79 [18,19,20,21]. Furthermore, the unique anatomical characteristics of the peritoneum pose a significant challenge in distinguishing it from surrounding visceral fat on imaging, thereby necessitating a higher level of expertise from the radiologist.
Radiomics (Rad) is a promising image data analysis method that has emerged in recent years, capable of quantifying traditional medical images into structured data features and interpreting images from multiple dimensions [22]. It has demonstrated potential in determining the benignity and malignancy of tumors, staging, and prognosis [23,24,25]. Similarly, deep learning (DL) can also extract quantified features from images with higher dimensionality and more accurate performance than Rad features (RF) [26]. Previous studies have shown that DL exhibits better performance than RF in predicting preoperative occult PM or lymphatic metastasis in gastric cancer patients [27, 28]. However, there are currently no reported studies utilizing DL for the preoperative staging of PM patients.
In summary, the purpose of this study was to develop and validate a DL model for the preoperative staging of PM patients to individualize patient treatment based on accurate staging.
Materials and methods
Patient enrolment
This study included 362 PM patients who underwent surgery at Zhongnan Hospital of Wuhan University from January 2014 to December 2021. The inclusion criteria for patients were as follows: (1) evaluation of PCI via laparoscopic exploration or open surgery, (2) confirmation of PM through postoperative pathological examination, and (3) receipt of a whole contrast-enhanced abdominal CT within three weeks prior to surgery. Patients were excluded if they had poor CT imaging quality or incomplete baseline clinical or pathological information. Of the initial pool of patients, 168 patients were included in the final analysis and categorized into two groups: Mild peritoneal involvement (MPI) group (PCI < 18) and Heavy peritoneal involvement (HPI) group (PCI ≥ 18), based on the S-PCI. This division considered the established indications for gastrointestinal CRS, which typically fall below the range of 16–20 for PCI. The training and validation cohorts were randomly divided into an 8:2 ratio. A flow chart of this study design is provided in Fig. 1.
The Medical Ethics Committee of Wuhan University Zhongnan Hospital approved this study. Since this study was retrospective, informed consent from patients was waived.
Intraoperative evaluation
The S-PCI was determined during laparoscopy exploration or open surgery by a team of surgeons. The abdominopelvic cavity was partitioned into 13 distinct regions, and the size of the tumor in each region was used to assign a score, with a score of 0 indicating no tumor, 1 for tumor diameter < 0.5 cm, 2 for tumor diameter ranging from 0.5 to 5 cm, and 3 for tumor diameter exceeding 5 cm. The summation of scores from all regions yielded the S-PCI, which could range from 0 to 39 [29]. All samples obtained during surgery were confirmed as PM through postoperative pathological examination.
CT image acquisition
CT examinations were conducted on 128-slice (Siemens SOMATOM Definition CT) and 64-channel (GE Discovery CT750, Philips Ingenuity CT) scanners with patients in the supine position following a 6-h fasting period. Patients were trained to control their breathing before the scan to minimize any breathing-related artifacts. The CT scans encompassed the region extending from the diaphragm to the bony pelvic floor. Prior to undergoing the contrast-enhanced CT examination, patients were administered contrast agents (5.300 mL/kg, iohexol 40 mg I/mL) via the anterior elbow vein at a rate of 5.1 mL/s. The following parameters were used for the CT scan: tube current ranging from 150 to 350 mA, tube voltage of 120 kVp, field of view spanning 30–45 cm, matrix size of 512 × 512, and reconstructed slice thickness between 1 and 5 mm.
Image preprocessing and segmentation
We chose CT images of the portal venous-phase for the relevant study because they could well separate between normal tissues, blood vessels, tumors and non-neoplastic organs. To reduce the effect of different volume pixels, we used the nearest interpolation method to resample the voxels to 1 mm × 1 mm × 1 mm. The 3D peritoneal volume of interest (VOI) segmentation was using ITK-SNAP software (v.4.0.0, http://www.itksnap.org). A novel approach to segmenting the 3D peritoneal volume of interest was developed in this study. This was achieved by dividing the entire abdomen into three sections: upper, middle, and lower abdomen. The abdomen was first segmented without organs and organs, musculoskeletal and vascular areas using semi-automatic segmentation on ITK-SNAP (see Supplementary 1 for details), and then by two experienced imaging physicians manually corrected the VOI in the abdominal window (window level: 30HU, window width: 700HU) to obtain the peritoneal VOI of the upper, middle and lower parts. Thirty patients were randomly selected after 30 days to have the VOI segmented again to exclude intra- and interobserver variation. The flow in the whole study was as in Fig. 2.
The workflow in this study. Firstly, the clinical and imaging data of the patients were collected. Subsequently, ITK-SNAP was employed to outline the imaging data and obtain the region of interest (ROI). The features of the ROI were then extracted using both radiomics and deep learning techniques. Finally, a machine learning-based peritoneal carcinomatosis index (PCI) prediction model was constructed based on filtered features. The direct predictive ability of different models was compared through evaluation metrics including area under the curve (AUC), confusion matrix, and decision curve analysis (DCA)
Radiomics feature extraction
RF extraction of peritoneal VOI was performed using PyRadiomics (v3.1.0) based on the Python (3.9.7) open-source platform. The types and distribution of RF areshown in the supplementary material. The extraction process aligns with the Imaging Biomarker Standardization Initiative (IBSI) [30]. The reproducibility of RF was assessed by the intraclass correlation coefficient (ICC), and features with ICC ≥ 0.75 proved to have low variance and high reproducibility.
Deep learning feature extraction
The layer with the largest peritoneal area from the three abdominal VOI of patients were selected as the patient’s DL region of interest (ROI). DenseNet, InceptionV3, ResNet18, ResNet50, and ResNet101 were pre-trained convolutional neural network models based on ImageNet. In the present study, DL models were utilized to extract DL features (DLF) from the ROI. To obtain the DLF, the last fully connected layer of the model was removed and global maximum pooling was applied to extract the maximum value of each feature layer. The original value of the output image features was used as the DLF. To visualize the DLF, the guided class activation map (Grad-CAM) was used. Grad-CAM is a visualization technique utilized to identify the key regions within an input image that significantly influenced the output of the neural network [31]. The PyTorch 1.4.1 library in Python 3.7.0 (https://pytorch.org) was used to implement the neural network.
Feature selection and model construction
Normalization of data is a common step in machine learning to scale and center the data, which helps in improving the accuracy of the model. In this study, the Z-Score method was used for normalization, which transforms the features to have a mean of zero and standard deviation of one. After normalization, Spearman correlation coefficient was calculated to identify highly correlated features, and one of the features with a strong correlation was retained (correlation coefficient > 0.9). To further select the most representative features, the least absolute shrinkage and selection operator (LASSO) regularization was introduced, which shrinks the regression coefficients towards zero and selects the features with non-zero coefficients. The regularization parameter λ was determined using five-fold cross-validation.
Nine machine learning methods (Table 2) were used to construct the classification model. The training and validation cohorts were randomly split in an 8:2 ratio, and the optimal model was selected based on the area under the receiver operating characteristic (ROC) curve (AUC). Finally, the performance of the model was evaluated using the decision curve analysis (DCA) and calibration curve.
Statistical analysis
Statistical analyses were performed using the Python 3.9.7 platform (https://www.python.org/). Continuous variables were presented as mean values and standard deviations, and analyzed using either Mann-Whitney U test or Student t-test. Categorical variables were reported as absolute number and analyzed using chi-square test or Fisher’s exact test. The Delong test was utilized to determine the statistical difference between the AUC values. A significance level of 0.05 was chosen for all tests, meaning that a p value below this level was considered statistically significant.
Results
Clinical characteristics of the patients
In this study, a total of 168 patients were recruited and allocated randomly into either the training cohort (n = 134) or the validation cohort (n = 34) at an 8:2 ratio. Table 1 presents the characteristics of all patients. Among the patients, 79 (47.02%) were male and 89 (52.98%) were female, while 89 (52.98%) belonged to the MPI group, and 79 (47.02%) were in the HPI group (Table 1). The distribution of each clinical information was balanced between the two cohorts. Chi-square and t tests indicated that Primary tumor, CA125, CA199, and L/M were significantly correlated with PCI in patients (P < 0.05). Subsequently, univariate and multivariate regression analyses were conducted, revealing that Primary tumor (P = 0.004), CA125 (P < 0.001), and L/M (P = 0.001) could serve as effective diagnostic indicators of PCI staging in patients preoperatively.
Feature extraction and selection for model construction
A total of 4683 RF were extracted from the upper, middle, and lower peritoneal ROIs, with 1561 features in each part. After excluding 1231 features with poor reproducibility (ICC < 0.75) based on ICC reproducibility assessment, 35 RF (as shown in Fig. 3A) were obtained for model construction through further screening with Spearman correlation coefficient and Lasso to exclude features with weak correlation. The Rad-SVM model demonstrated the best prediction performance (as demonstrated in Table 2), and the AUC in the validation cohort was 0.906 (95% CI 0.804–1.000).
Regarding the combination of six DL models and nine machine learning methods used for feature extraction, DenseNet121-SVM demonstrated the best performance (as shown in Table 3), with 4704 DLF extracted (1568 features each in the upper, middle, and lower abdomen). After filtering, 53 DLF were finally selected (as shown in Fig. 3B), and the AUC of the constructed model in the validation cohort was 0.951 (95% CI 0.887–1.000).
Performance of diagnostic models
We developed three pre-procedural diagnostic models for predicting PCI using clinical data, RF, and DLF. It is worth noting that the AUC of both Rad and DL models was significantly higher than Clinic model. In the training cohort, the AUC of DL model was 0.996, which was higher than Clinic model (AUC = 0.851, P < 0.001) and Rad model (AUC = 0.986, P = 0.1468). In the validation cohort, the AUC of DL model was 0.951, which was higher than Clinic model (AUC = 0.797, P = 0.074) and Rad model (AUC = 0.906, P = 0.4654), as shown in Fig. 4A and D.
The results of the DCA demonstrated that patients could derive greater net benefits from DL model in comparison to both Clinic and Rad models (Fig. 4B and E). Additionally, the calibration curves indicated a superior level of agreement between predicted and actual groupings within DL models (Fig. 4C, F).
We also constructed a Clinic-Rad-DL model, which had an AUC of 0.962 (95% CI 0.901–1.000) in the validation cohort. However, the Clinic-Rad-DL model did not result in a substantial improvement in efficacy compared to DL model alone (P = 0.695).
While the features extracted by DL may be challenging to interpret in practical terms, the study implemented Grad-CAM to provide interpretability to the DL model. This involved characterizing the distribution of contributions to the output prediction results (Fig. 5).
Furthermore, we assessed the generalization ability of the DL model in practical applications by dividing patients into different subgroups. The DL model showed strong generalization ability in the stratified analysis according to age, gender, primary tumor, presence or absence of ascites, and different CT layer thicknesses (Table 4).
Discussion
The purpose of this study was to develop and validate a DL model using preoperative enhanced abdominal CT to determine the degree of preoperative peritoneal involvement in PM patients. The results demonstrated that the model had satisfactory performance in stratifying patients and could potentially aid in treatment decision-making for PM patients.
Clinically, PM patients are usually at the end stage of the disease. Among the limited treatment options, CRS + HIPEC can maximize the overall survival time of patients, but this is only for patients with a small PM load [32, 33]. However, accurate preoperative staging of patients can avoid unnecessary surgical damage. Imaging examination, the most commonly used noninvasive diagnostic modality, has a limited ability to identify PM [34, 35]. CT combined with DL analysis in patients with PM has been widely used in predicting occult metastasis and recurrence of PM, demonstrating a strong predictive capability [27, 36, 37]. In a study of 163 patients, Zhang and colleagues [38] attempted to use radiomics rather than DL for preoperative PCI scoring, and used an external validation cohort to assess the applicability of the model. In their study, limited by intact peritoneal segmentation, only 6 regions were selected as representatives. In our study, we are the first to segment the complete peritoneum and use DL analysis of CT images for staging patients with preoperative PM. The results demonstrate the superior discriminatory performance of the Rad and DL models over clinical data alone. The DL model demonstrated superior AUC performance compared to the Rad model in both training and validation cohorts, while the DCA curve indicated that both models could enhance patient outcomes following CRS + HIPEC treatment. Additionally, Grad-CAM was used to address the black box problem of DL and provide clinicians with explanations. The focus on mesenteric heterogeneity in DL is consistent with what clinicians see intraoperatively, as the mesentery is a common site for PM.
In previous studies on using Rad or DL for PM, one of the main challenges was the segmentation of the ROI. Manual segmentation of ROI is subjective and can vary among different imaging physicians. Furthermore, the complexity of the peritoneum makes it difficult to manually sketch the entire peritoneum, and most studies [3, 27, 36, 39] only assess the status of the peritoneum in primary foci ROI or a portion of the peritoneum, which may not represent the entire peritoneum. Additionally, there is significant variation in the primary foci in patients with PM from numerous sources, and using only the primary foci ROI does not generalize the entire model to all patients with PM of tumor origin. To address these issues, we segmented the entire peritoneal region of patients semi-automatically using ITK-SNAP to segment the 3D peritoneal VOI for the first time. This approach allows us to accurately and comprehensively assess the peritoneal status of PM patients efficiently. Moreover, we divided the whole abdomen into three parts, and found that the majority of PM patients of gastric origin in our cohort and the majority of the finally screened RF and DLF also originated from the upper and middle abdomen. Clinically, gastric cancer PM is rarely found in the lower abdomen. This finding suggests that we can select different abdominal partitions for the classification of primary tumor origin to further optimize the efficacy of the model. Overall, the semi-automatic segmentation of the peritoneal VOI provides a more objective and comprehensive assessment of PM patients, and dividing the abdomen into different partitions may improve the model’s efficacy in identifying the primary tumor origin.
In our study cohort, a significant proportion of patients (47.0%) with S-PCI scores of 18 or higher underwent surgical exploration but did not receive any further surgical treatment, indicating that our DL model can potentially optimize available surgical options. Furthermore, our model holds promise for application in patients with higher PM load, where multi-stage HIPEC may be a viable option for achieving surgical indications similar to preoperative neoadjuvant chemotherapy [3]. As such, we anticipate that our model can enable selection of patients for surgical treatment and evaluation of HIPEC-treated SIP patients in the future.
However, there are limitations to our study that must be considered. First, fewer patients were evaluated for the S-PCI procedure, and larger samples may be needed to improve the reliability. Second, as our study was retrospective and based on a single-center cohort, the generalizability of the model must be validated with prospective and multicenter data. Finally, while DL analysis in medical imaging is primarily based on 2D imaging, the potential for future models utilizing 3D imaging analysis has the capacity to further optimize model performance.
In summary, our study presents a promising CT-based DL model for preoperative staging assessment of PM patients, providing a non-invasive means to optimize clinical treatment decisions.
References
Sadeghi B, Arvieux C, Glehen O et al (2000) Peritoneal carcinomatosis from non-gynecologic malignancies: results of the EVOCAPE 1 multicentric prospective study. Cancer 88(2):358–363
Gwee YX, Chia D, So J et al (2022) Integration of genomic biology into therapeutic strategies of gastric cancer peritoneal metastasis. J Clin Oncol 40(24):2830
van Stein RM, Aalbers A, Sonke GS, van Driel WJ (2021) Hyperthermic intraperitoneal chemotherapy for ovarian and colorectal cancer: a review. JAMA Oncol 7(8):1231–1238
Ji ZH, Peng KW, Yu Y et al (2017) Current status and future prospects of clinical trials on CRS + HIPEC for gastric cancer peritoneal metastases. Int J Hyperthermia 33(5):562–570
Li Y, Zhou YF, Liang H et al (2016) Chinese expert consensus on cytoreductive surgery and hyperthermic intraperitoneal chemotherapy for peritoneal malignancies. World J Gastroenterol 22(30):6906–6916
Bonnot PE, Piessen G, Kepenekian V et al (2019) Cytoreductive surgery with or without hyperthermic intraperitoneal chemotherapy for gastric cancer with peritoneal metastases (CYTO-CHIP study): a propensity score analysis. J Clin Oncol 37(23):2028–2040
Esquivel J, Lowy AM, Markman M et al (2014) The American Society of Peritoneal Surface Malignancies (ASPSM) multiinstitution evaluation of the peritoneal surface disease severity score (PSDSS) in 1,013 patients with colorectal cancer with peritoneal carcinomatosis. Ann Surg Oncol 21(13):4195–4201
van Driel WJ, Koole SN, Sikorska K et al (2018) Hyperthermic intraperitoneal chemotherapy in ovarian cancer. N Engl J Med 378(3):230–240
Schwarz L, Votanopoulos K, Morris D et al (2016) Is the combination of distal pancreatectomy and cytoreductive surgery with HIPEC reasonable? Results of an international multicenter study. Ann Surg 263(2):369–375
Amblard I, Mercier F, Bartlett DL et al (2018) Cytoreductive surgery and HIPEC improve survival compared to palliative chemotherapy for biliary carcinoma with peritoneal metastasis: a multi-institutional cohort from PSOGI and BIG RENAPE groups. Eur J Surg Oncol 44(9):1378–1383
Bonnot PE, Lintis A, Mercier F et al (2021) Prognosis of poorly cohesive gastric cancer after complete cytoreductive surgery with or without hyperthermic intraperitoneal chemotherapy (CYTO-CHIP study). Br J Surg 108(10):1225–1235
Sugarbaker PH (2016) Cytoreductive surgery and hyperthermic intraperitoneal chemotherapy in the management of gastrointestinal cancers with peritoneal metastases: progress toward a new standard of care. Cancer Treat Rev 48:42–49
Lorimier G, Linot B, Paillocher N et al (2017) Curative cytoreductive surgery followed by hyperthermic intraperitoneal chemotherapy in patients with peritoneal carcinomatosis and synchronous resectable liver metastases arising from colorectal cancer. Eur J Surg Oncol 43(1):150–158
Elias D, Gilly F, Boutitie F et al (2010) Peritoneal colorectal carcinomatosis treated with surgery and perioperative intraperitoneal chemotherapy: retrospective analysis of 523 patients from a multicentric french study. J Clin Oncol 28(1):63–68
Ji ZH, Yu Y, Liu G et al (2021) Peritoneal cancer index (PCI) based patient selecting strategy for complete cytoreductive surgery plus hyperthermic intraperitoneal chemotherapy in gastric cancer with peritoneal metastasis: a single-center retrospective analysis of 125 patients. Eur J Surg Oncol 47(6):1411–1419
Jacquet P, Sugarbaker PH (1996) Clinical research methodologies in diagnosis and staging of patients with peritoneal carcinomatosis. Cancer Treat Res 82:359–374
van Oudheusden TR, Braam HJ, Luyer MD et al (2015) Peritoneal cancer patients not suitable for cytoreductive surgery and HIPEC during explorative surgery: risk factors, treatment options, and prognosis. Ann Surg Oncol 22(4):1236–1242
Van’t Sant I, van Eden WJ, Engbersen MP et al (2019) Diffusion-weighted MRI assessment of the peritoneal cancer index before cytoreductive surgery. Br J Surg 106(4):491–498
Dohan A, Hoeffel C, Soyer P et al (2017) Evaluation of the peritoneal carcinomatosis index with CT and MRI. Br J Surg 104(9):1244–1249
Low RN, Barone RM, Lucero J (2015) Comparison of MRI and CT for predicting the peritoneal cancer index (PCI) preoperatively in patients being considered for cytoreductive surgical procedures. Ann Surg Oncol 22(5):1708–1715
Lin CN, Huang WS, Huang TH et al (2021) Adding value of MRI over CT in predicting peritoneal cancer index and completeness of cytoreduction. Diagnostics (Basel) 11(4):674
Mayerhoefer ME, Materka A, Langs G et al (2020) Introduction to radiomics. J Nucl Med 61(4):488–495
Xu X, Zhang HL, Liu QP et al (2019) Radiomic analysis of contrast-enhanced CT predicts microvascular invasion and outcome in hepatocellular carcinoma. J Hepatol 70(6):1133–1144
Basso Dias A, Mirshahvalad SA, Ortega C et al (2023) The role of [(18)F]-DCFPyL PET/MRI radiomics for pathological grade group prediction in prostate cancer. Eur J Nucl Med Mol Imaging 50:2167–2176
Yang L, Gu D, Wei J et al (2019) A radiomics nomogram for preoperative prediction of microvascular invasion in hepatocellular carcinoma. Liver Cancer 8(5):373–386
Bera K, Braman N, Gupta A, Velcheti V, Madabhushi A (2022) Predicting cancer outcomes with radiomics and artificial intelligence in radiology. Nat Rev Clin Oncol 19(2):132–146
Jiang Y, Liang X, Wang W et al (2021) Noninvasive prediction of occult peritoneal metastasis in gastric cancer using deep learning. JAMA Netw Open 4(1):e2032269
Dong D, Fang MJ, Tang L et al (2020) Deep learning radiomic nomogram can predict the number of lymph node metastasis in locally advanced gastric cancer: an international multicenter study. Ann Oncol 31(7):912–920
Ihemelandu C (2023) The landmark series: scoring systems for primary peritoneal surface malignancy. Ann Surg Oncol 30(3):1832–1837
Zwanenburg A, Vallières M, Abdalah MA et al (2020) The image biomarker standardization initiative: standardized quantitative radiomics for high-throughput image-based phenotyping. Radiology 295(2):328–338
Selvaraju RR, Cogswell M (2019) Grad-CAM: visual explanations from deep networks via gradient-based localization. Preprint at https://arxiv.org/abs/1610.02391
Granieri S, Bonomi A, Frassini S et al (2021) Prognostic impact of cytoreductive surgery (CRS) with hyperthermic intraperitoneal chemotherapy (HIPEC) in gastric cancer patients: a meta-analysis of randomized controlled trials. Eur J Surg Oncol 47(11):2757–2767
Wong L, Li Z, Fan Q et al (2022) Cytoreductive surgery (CRS) with hyperthermic intraperitoneal chemotherapy (HIPEC) in peritoneal sarcomatosis—a systematic review and meta-analysis. Eur J Surg Oncol 48(3):640–648
Chen J, Liu S, Tang Y et al (2020) Diagnostic performance of CT for differentiating peritoneal tuberculosis from peritoneal carcinomatosis: a systematic review and meta-analysis. Clin Radiol 75(5):396e7-396e14
Chang MC, Chen JH, Liang JA, Huang WS, Cheng KY, Kao CH (2013) PET or PET/CT for detection of peritoneal carcinomatosis: a meta-analysis. Clin Nucl Med 38(8):623–629
Jiang Y, Zhang Z, Yuan Q et al (2022) Predicting peritoneal recurrence and disease-free survival from CT images in gastric cancer with multitask deep learning: a retrospective study. Lancet Digit Health 4(5):e340–e350
Yuan Z, Xu T, Cai J et al (2022) Development and validation of an image-based deep learning algorithm for detection of synchronous peritoneal carcinomatosis in colorectal cancer. Ann Surg 275(4):e645–e651
Zhang Q, Yuan Y, Li S et al (2022) A CT-based radiomics model for evaluating peritoneal cancer index in peritoneal metastasis cases: a preliminary study. Acad Radiol 30(7):1329–1339
Mirniaharikandehei S, Heidari M, Danala G, Lakshmivarahan S, Zheng B (2021) Applying a random projection algorithm to optimize machine learning model for predicting peritoneal metastasis in gastric cancer patients using CT images. Comput Methods Programs Biomed 200:105937
Funding
This work was supported by grants from the Improvement Project for Theranostic ability on Difficulty miscellaneous disease (Tumor) (No. ZLYNXM202018).
Author information
Authors and Affiliations
Contributions
Conceptualization, JW; Data curation, YH, HX, and TS; Funding acquisition, BX; Methodology, JW and YH. Project administration, HX and BX; Resources, YH and TS; Supervision, SW and BX; Validation, JW; Writing – original draft, JW; Writing – review & editing, YH, SW, HX and BX.
Corresponding authors
Ethics declarations
Conflict of interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Wang, J., Hu, Y., Xiong, H. et al. CT-based deep learning model: a novel approach to the preoperative staging in patients with peritoneal metastasis. Clin Exp Metastasis 40, 493–504 (2023). https://doi.org/10.1007/s10585-023-10235-5
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10585-023-10235-5