Computed tomography-based radiomics diagnostic approach for differential diagnosis between early- and late-stage pancreatic ductal adenocarcinoma

Abstract

Background: One of the primary reasons for the dismal survival rates in pancreatic ductal adenocarcinoma (PDAC) is that most patients are usually diagnosed at late stages. There is an urgent unmet clinical need to identify and develop diagnostic methods that could precisely detect PDAC at its earliest stages.

Aim: To evaluate the potential value of radiomics analysis in the differentiation of early-stage PDAC from late-stage PDAC.

Methods: A total of 71 patients with pathologically proved PDAC based on surgical resection who underwent contrast-enhanced computed tomography (CT) within 30 d prior to surgery were included in the study. Tumor staging was performed in accordance with the 8^th edition of the American Joint Committee on Cancer staging system. Radiomics features were extracted from the region of interest (ROI) for each patient using Analysis Kit software. The most important and predictive radiomics features were selected using Mann-Whitney U test, univariate logistic regression analysis, and minimum redundancy maximum relevance (MRMR) method. Random forest (RF) method was used to construct the radiomics model, and 10-times leave group out cross-validation (LGOCV) method was used to validate the robustness and reproducibility of the model.

Results: A total of 792 radiomics features (396 from late arterial phase and 396 from portal venous phase) were extracted from the ROI for each patient using Analysis Kit software. Nine most important and predictive features were selected using Mann-Whitney U test, univariate logistic regression analysis, and MRMR method. RF method was used to construct the radiomics model with the nine most predictive radiomics features, which showed a high discriminative ability with 97.7% accuracy, 97.6% sensitivity, 97.8% specificity, 98.4% positive predictive value, and 96.8% negative predictive value. The radiomics model was proved to be robust and reproducible using 10-times LGOCV method with an average area under the curve of 0.75 by the average performance of the 10 newly built models.

Conclusion: The radiomics model based on CT could serve as a promising non-invasive method in differential diagnosis between early and late stage PDAC.

Keywords: American Joint Committee on Cancer staging; Computed tomography; Pancreatic ductal adenocarcinoma; Radiomics.

Figure 1

Flow diagram of patient inclusion. PDAC: Pancreatic ductal adenocarcinoma; CE-CT: Contrast-enhanced computed tomography; AJCC: American Joint Committee on Cancer.

Figure 2

Computed tomography images of two representative cases of pancreatic ductal adenocarcinoma. A-C: A 54-year-old man with stage IB pancreatic ductal adenocarcinoma (PDAC). Computed tomography (CT) imaging showed a well-defined, predominantly solid tumor (white arrow), 3.0 cm in diameter in the tail of the pancreas. No obvious lymph node metastasis, vascular invasion, or distal metastasis was found. TNM classification: T2N0M0 (IB); D-F: A 44-year-old man with stage IV PDAC. CT imaging showed an ill-defined, predominantly solid tumor (white arrow), 4.1 cm in diameter in the tail of the pancreas. The left adrenal gland (orange arrow) was grossly invaded by the tumor. TNM classification: T3N1M1 (IV).

Figure 3

Histograms of two representative cases of pancreatic ductal adenocarcinoma. A and B: Stage IB pancreatic ductal adenocarcinoma (PDAC), A vs stage IV PDAC, B, which showed a marked difference. The X-axis indicates the gray level (HU). The Y-axis indicates the number of voxels.

Figure 4

The area under receiver operating characteristic curve barplot, importance, and heatmap showing expression values of the selected features. The feature importance plot was provided by the random forest methods as soon as the model train procedure completed. The importance of the feature was computed from permuting out-of-bag data. A: The area under receiver operating characteristic curve barplot of the selected predictive radiomics features; B: The importance of the selected predictive radiomics features; C: The heatmap showing expression values of the selected radiomics features (rows) of 71 pancreatic ductal adenocarcinoma patients (columns). AUC: Area under the receiver operating characteristic curve.

Figure 5

Receiver operating characteristic curve of the radiomics model and receiver operating characteristic curves of 10-times leave group out cross-validation analysis in differential diagnosis between early and late-stage pancreatic ductal adenocarcinoma. A: Receiver operating characteristic (ROC) curve of the radiomics model based on computed tomography images in the differentiation of early and late-stage pancreatic ductal adenocarcinoma (PDAC); B: ROC curves of 10-times leave group out cross-validation analysis for the differentiation of early and late stage PDAC, with a mean area under the curve of 0.75 by the average performance of the 10 newly built models, indicating good predictive ability of the radiomics model. AUC: Area under the receiver operating characteristic curve.