Contrast-enhanced CT radiomics combined with multiple machine learning algorithms for preoperative identification of lymph node metastasis in pancreatic ductal adenocarcinoma

Abstract

Objectives: This research aimed to assess the value of radiomics combined with multiple machine learning algorithms in the diagnosis of pancreatic ductal adenocarcinoma (PDAC) lymph node (LN) metastasis, which is expected to provide clinical treatment strategies.

Methods: A total of 128 patients with pathologically confirmed PDAC and who underwent surgical resection were randomized into training (n=93) and validation (n=35) groups. This study incorporated a total of 13 distinct machine learning algorithms and explored 85 unique combinations of these algorithms. The area under the curve (AUC) of each model was computed. The model with the highest mean AUC was selected as the best model which was selected to determine the radiomics score (Radscore). The clinical factors were examined by the univariate and multivariate analysis, which allowed for the identification of factors suitable for clinical modeling. The multivariate logistic regression was used to create a combined model using Radscore and clinical variables. The diagnostic performance was assessed by receiver operating characteristic curves, calibration curves, and decision curve analysis (DCA).

Results: Among the 233 models constructed using arterial phase (AP), venous phase (VP), and AP+VP radiomics features, the model built by applying AP+VP radiomics features and a combination of Lasso+Logistic algorithm had the highest mean AUC. A clinical model was eventually constructed using CA199 and tumor size. The combined model consisted of AP+VP-Radscore and two clinical factors that showed the best diagnostic efficiency in the training (AUC = 0.920) and validation (AUC = 0.866) cohorts. Regarding preoperative diagnosis of LN metastasis, the calibration curve and DCA demonstrated that the combined model had a good consistency and greatest net benefit.

Conclusions: Combining radiomics and machine learning algorithms demonstrated the potential for identifying the LN metastasis of PDAC. As a non-invasive and efficient preoperative prediction tool, it can be beneficial for decision-making in clinical practice.

Keywords: computed tomography; lymph node metastasis; machine learning; pancreatic ductal adenocarcinoma; radiomics.

Figure 1

Flow chart of inclusion and exclusion criteria for patients.

Figure 2

The workflow of model construction and validation.

Figure 3

Develop and validate models through integrated machine learning. (A) A total of 67 prediction models were built using features from arterial phase radiomics, and the AUC of each model was further computed on the training and validation datasets. (B) 84 predictive models built from venous phase radiomics features. (C) 82 predictive models constructed based on radiomics features of arterial and venous phases. (D) Wayne diagram of the number of models.

Figure 4

Radiomics features selection. (A) The value (λ) was chosen for the Lasso algorithm’s tuning parameter through 10-fold cross-validation. Two vertical lines show the optimal values according to the minimum criterion and 1-SE criterion. The optimal λ value of 0.0229 was chosen. (B) Profiles of Lasso coefficients for 41 radiomics features. A vertical line was drawn at the value selected using 10-fold cross-validation, and a total of nine features with non-zero coefficients were chosen. (C) Nine selected radiomics features and their coefficients.

Figure 5

The comparative predictive performance of various models. ROC curves of different models in the training (A) and validation (B) cohorts; Heat map of the P-values for the DeLong test between the training (C) and validation (D) cohorts of different models.

Figure 6

(A) A nomogram was plotted by combining AP+VP-Radscore with independent clinical predictors in the training cohort. Nomogram calibration curves for the training (B) and validation (C) cohorts.

Figure 7

DCA curves for three models in the training (A) and validation sets (B).