Prediction of regional lymph node status in rectal cancer with radiomics features based on deep learning segmented tumor area

Abstract

Background: To predict regional lymph node metastasis (LNM) in rectal cancer (RC) using deep learning-based tumor auto-segmentation and radiomics. Methods: This single-center research retrospectively analyzed 282 patients with RC from two MR vendors. The deep learning-based auto-segmentation models were constructed on T2WI and DWI with 3D U-Net, 3D V-Net, and nnU-Net v2 and assessed with the Dice Similarity Coefficient (DSC). Radiomics features on manual-based VOI (MbV) and deep learning-based VOI (DbV, with the highest DSC) were extracted respectively. After feature normalization and selection, five classifiers were used for radiomics model building and then for LNM prediction. The optimal model was selected using a 5-fold cross-validation strategy and evaluated with area under the curve (AUC), accuracy, specificity, and sensitivity. Results: The DSC of the nnU-Net v2 was significantly higher than that of the 3D U-Net and 3D V-Net (T2WI: 0.886 vs 0.548 vs 0.616, p < 0.001; DWI: 0.906 vs 0.583 vs 0.433, p < 0.001; test set). The AUC of DbV based-radiomics models (0.700 for T2WI, 0.667 for DWI, and 0.800 for T2WI + DWI) were comparable to those of the corresponding MbV-based radiomics models (0.633 for T2WI, p = 0.638; 0.700 for DWI, p = 0.544; and 0.833 for T2WI + DWI, p = 0.248) in LNM prediction. Conclusions: Radiomics features of T2WI and DWI based on nnU-net v2 segmented tumor area showed a reliable performance in predicting LNM in RC.