Development and Validation of a Deep Learning Radiomics Model Combined with Clinical Radiological Features for Predicting Occult Peritoneal Metastasis in Patients with Pancreatic Ductal Adenocarcinoma

Background

Pancreatic ductal adenocarcinoma (PDAC) is an extremely lethal malignancy with a 5-year survival rate of approximately 11%. The poor prognosis is partly due to the fact that 80-85% of patients are diagnosed with advanced, unresectable, or metastatic disease, including occult peritoneal metastases (OPM), when they become symptomatic. The peritoneum is the second most common site of PDAC metastasis, with approximately 10-20% of patients presenting with peritoneal metastasis at initial diagnosis. For these patients, early identification of peritoneal metastasis will greatly influence the choice of treatment and avoid unnecessary surgery.

Conventional diagnosis of peritoneal metastasis relies on computed tomography (CT), but early peritoneal metastasis is often difficult to detect due to the lack of obvious signs. Diagnostic staging laparoscopy, although recommended for the diagnosis of peritoneal metastasis, is invasive and not the most cost-effective choice. Some clinical and radiological features, such as cancer antigen 19-9 (CA19-9) and regional lymph node enlargement, are considered predictors of peritoneal metastasis, but their predictive accuracy has not been widely confirmed. Therefore, there is an urgent need to develop a personalized model to predict peritoneal metastasis in PDAC patients before treatment.

In recent years, radiomics and deep learning techniques have achieved significant results in cancer prediction. However, there is currently no deep learning radiomics model specifically for OPM in PDAC patients. Therefore, this study aimed to develop and validate a model that combines handcrafted radiomics (HCR) and deep learning radiomics (DLR) to predict OPM in PDAC patients.

Study Source

This study was conducted by multiple authors including Siya Shi, Chuxuan Lin, Jian Zhou, and others, affiliated with institutions such as the First Affiliated Hospital of Sun Yat-sen University in Guangzhou, and the School of Medicine at Shenzhen University. The paper was published in the International Journal of Surgery on March 4, 2024.

Research Process

Study Subjects and Grouping

This was a retrospective, dual-center study that included 302 PDAC patients who underwent baseline CT scans between January 2012 and October 2022. The study was divided into a training cohort (167 patients), internal test cohort (72 patients), and external test cohort (63 patients). In each cohort, patients were labeled as either OPM-positive or OPM-negative (e.g., 22 patients were OPM-positive in the training cohort).

Data Collection and Processing

CT Scanning Protocol: Various CT scanners from Canon Medical Systems and Philips Healthcare were used. Detailed CT scanning protocols and parameters are provided in the supplementary materials.

Data Extraction: The study collected clinical and radiological parameters for all patients, including gender, age, tumor location, serum carcinoembryonic antigen (CEA) and CA19-9 levels, and CT-based T and N staging.

Feature Extraction and Selection: Three-dimensional and two-dimensional segmentation was performed using the MITK software, and 1130 tumor HCR features, 512 tumor DLR features, and 474 peritoneal HCR features were extracted. For feature selection, mutual information and the least absolute shrinkage and selection operator (LASSO) algorithm were first used for screening, followed by modeling with a logistic regression model.

Model Development and Validation

Four models were developed in this study: a clinical-radiological model, a tumor DLR model, a peritoneal HCR model, and a combined model. The performance of each model was evaluated using the receiver operating characteristic (ROC) curve and the area under the curve (AUC). The process involved feature selection using the LASSO algorithm in the training cohort, and validation in the internal and external test cohorts.

Study Results

Feature Selection and Model Performance: After feature selection, the combined model included 9 tumor HCR features, 14 tumor DLR features, 3 peritoneal HCR features, and 3 clinical-radiological features (CA19-9, CT-based T and N staging). With these features, the model achieved AUC values of 0.853 in the training cohort, 0.845 in the internal test cohort, and 0.852 in the external test cohort, demonstrating good predictive performance. Additionally, the DeLong test showed that the combined model was significantly better than the model containing only clinical-radiological features.

Model Comparison and Clinical Assessment: The combined model, which incorporated DLR and clinical-radiological features, had higher overall net benefit in both the training and test cohorts compared to using clinical-radiological features or other models alone. Decision curve analysis also indicated that the combined model had better clinical applicability at most reasonable thresholds.

Study Conclusions

The developed model, combining CT image DLR features and clinical-radiological data, demonstrated good predictive performance for OPM in PDAC patients. This non-invasive predictive tool can not only help identify OPM-positive patients to avoid unnecessary surgery but also screen for patients suitable for diagnostic laparoscopy and neoadjuvant therapy, providing additional evidence for personalized diagnosis and treatment.

Significance and Value of the Study

This study is the first to apply deep learning radiomics to the prediction of occult peritoneal metastasis in PDAC patients, demonstrating good performance in both the training and external validation cohorts. These research results indicate that by combining deep learning algorithms and clinical data, accurate prediction of occult peritoneal metastasis in PDAC patients can be achieved, providing important complementary value for clinical practice.

Research Highlights

1. This is the first dual-center deep learning radiomics study for occult peritoneal metastasis in PDAC patients, demonstrating good predictive performance and generalization ability.
2. By combining deep learning algorithms and clinical-radiological data, the model’s predictive accuracy was significantly improved, avoiding overfitting issues common in traditional models.
3. This non-invasive predictive tool can not only assist in diagnosis but also provide important clinical guidance for personalized treatment, helping to optimize patient treatment plans and prognosis.

This study showcases the great potential of radiomics in tumor diagnosis and treatment and provides a solid foundation for future clinical applications and further research.