Development and Validation of a Deep Learning Radiomics Model Predicting Lymph Node Status in Operable Cervical Cancer

Abstract

Aim: To develop and validate a deep learning radiomics model, which could predict the lymph node metastases preoperatively in cervical cancer patients. Patients and methods: We included a cohort of 226 pathological proven operable cervical cancer patients in two academic medical institutions from December 2014 to November 2017. Then this dataset was split into training set (n = 176) and independent testing set (n = 50) randomly. Five radiomic features were selected and a radiomic signature was established. We then combined these five radiomic features with the preoperative tumor histology and grade of these patients together. Baseline logistic regression model (LRM) and support vector machine model (SVM) were established for the comparison. We then explored the performance of a deep neural network (DNN), which is a popular deep learning model nowadays. Finally, performance of this DNN was validated in another independent test set including 50 cases of operable cervical cancer patients. Results: One thousand forty-five radiomic features were extracted for each patient. Twenty-eight features were found to be significantly correlated with the lymph node status in these patients (P < 0.05). Five radiomic features were further selected for further study due to their higher predictive powers. Baseline LRM incorporating these five radiomic and two clinicopathological features was established, which had an area under receiver operating characteristic curve (ROC) of 0.7372 and an accuracy of 89.20%. The established DNN model had four neural layers, in which layer there were 10 neurons. Adagrad optimizer and 1,500 iterations were used in training. The trained DNN had an area under curve (AUC) of 0.99 and an accuracy of 97.16% in the internal validation. To exclude the overfitting, independent external validation was also performed. AUC and accuracy in test set could still retain 0.90 and 92.00% respectively. Conclusion: This study used deep learning method to provide a comprehensive predictive model using preoperative CT images, tumor histology, and grade in cervical cancer patients. This model showed an acceptable accuracy in the prediction of lymph node status in cervical cancer. Our model may help identifying those patients who could benefit a lot from radiation therapy rather than primary hysterectomy surgery if this model could resist strict testing of future randomized controlled trials (RCTs).

Keywords: cervical cancer; deep learning; deep neural network; lymph node status; radiomics.

Figure 1

Global analysis pipeline.

Figure 2

Delineation of the primary tumor on the CT images of cervical cancer patients. Two expert radiologists performed the contouring. (A) The CT images of primary cancer in patient 1. (B) The contouring of the primary cancer for patient 1. (C) Tumor of the patient 2 on the CT images. (D) Contouring of the cancer on images of patient 2.

Figure 3

ROC of radiomic signature. Five radiomic features were selected due to their higher C index values (P < 0.5).

Figure 4

ROC of a baseline logistic regression predictive model incorporating the radiomic and clinicopathological features.

Figure 5

ROC of another baseline support vector machine model.

Figure 6

Results of cost function in the DNN model were decreased after each iteration. Learning rate was chosen as 0.1.

Figure 7

ROC of the trained DNN model after 1,500 iterations in the internal validation.

Figure 8

ROC of the trained DNN model tested in an independent set including 50 cases of operable cervical cancer.