An MRI-Based Deep Learning Radiomics Nomogram Analysis for Prediction of Postoperative Lymph Node Metastasis in Patients with Locally Advanced Cervical Cancer Receiving Neoadjuvant Therapy

Abstract

Abstract Objective The aim of this study was to investigate the value of combining radiomics features and deep learning features to construct Nomogram for non-invasive prediction of postoperative lymph node metastasis (LNM) by magnetic resonance imaging (MRI) prior to neoadjuvant chemotherapy (NACT) for cervical cancer, to assist clinical decision-making and diagnosis. Method Two hundred and sixty-five cervical cancer patients were divided into training (n = 212) and test (n = 53) sets in an 8:2 ratio. Radiomics features were extracted from Axial Liver Acquisition with Volume Acceleration plus Contrast enhancement (Ax-LAVA + C) sequences of MRI, and deep learning (DL) features were extracted using the Inception V3 model. The features selected by LASSO were combined with four machine learning algorithms to construct models to evaluate the predictive value of the radiomics features and DL features for postoperative LNM in cervical cancer patients operated after NACT. A Nomogram was constructed based on Logistic Regression model combining radiomics score (rad-score) and DL-score. Results In the radiomics model, the Multilayer Perceptron (MLP) outperforms other machine learning algorithms, with AUCs of 0.79,0.80 in the training set and test set, respectively. In the DL model, Support Vector Machine (SVM) outperforms other machine learning algorithms with AUCs of 0.78,0.78 in the training set and test set, respectively. The Nomogram constructed by combining radiomics features and deep learning features has an AUC of 0.93,0.89 in the training set and test set, respectively. It outperforms the radiomics model and the DL model. The decision curve analysis (DCA) shows that the Nomogram has good clinical benefits. Conclusion Radiomics models and deep learning models can effectively predict LNM status after NACT in cervical cancer patients. The Nomogram constructed by combining radiomics and DL features has better prediction performance compared with the radiomics model and DL model.