Automated Prediction of Neoadjuvant Chemoradiotherapy Response in Locally Advanced Cervical Cancer Using Hybrid Model-Based MRI Radiomics

Abstract

Background: This study aimed to develop a model that automatically predicts the neoadjuvant chemoradiotherapy (nCRT) response for patients with locally advanced cervical cancer (LACC) based on T2-weighted MR images and clinical parameters.

Methods: A total of 138 patients were enrolled, and T2-weighted MR images and clinical information of the patients before treatment were collected. Clinical information included age, stage, pathological type, squamous cell carcinoma (SCC) level, and lymph node status. A hybrid model extracted the domain-specific features from the computational radiomics system, the abstract features from the deep learning network, and the clinical parameters. Then, it employed an ensemble learning classifier weighted by logistic regression (LR) classifier, support vector machine (SVM) classifier, K-Nearest Neighbor (KNN) classifier, and Bayesian classifier to predict the pathologic complete response (pCR). The area under the receiver operating characteristics curve (AUC), accuracy (ACC), true positive rate (TPR), true negative rate (TNR), and precision were used as evaluation metrics.

Results: Among the 138 LACC patients, 74 were in the pCR group, and 64 were in the non-pCR group. There was no significant difference between the two cohorts in terms of tumor diameter (p = 0.787), lymph node (p = 0.068), and stage before radiotherapy (p = 0.846), respectively. The 109-dimension domain features and 1472-dimension abstract features from MRI images were used to form a hybrid model. The average AUC, ACC, TPR, TNR, and precision of the proposed hybrid model were about 0.80, 0.71, 0.75, 0.66, and 0.71, while the AUC values of using clinical parameters, domain-specific features, and abstract features alone were 0.61, 0.67 and 0.76, respectively. The AUC value of the model without an ensemble learning classifier was 0.76.

Conclusions: The proposed hybrid model can predict the radiotherapy response of patients with LACC, which might help radiation oncologists create personalized treatment plans for patients.

Keywords: MRI radiomics; automated prediction; cervical cancer; radiotherapy response.

Figure 1

MRI images of a patient with pCR before radiotherapy (left) and a patient with non-pCR before radiotherapy (right). We used T2WI-FS here to provide a clearer image of the collected data; The red circles represent the tumor area.

Figure 2

The pipeline of the proposed method.

Figure 3

The structure of the VGG19 network.

Figure 4

Ensemble learning flow charts for pCR/non-pCR prediction.

Figure 5

Maximum response graphs of convolution kernel in (a) Block1-2, (b) Block2-2, (c) Block3-4, (d) Block4-4, and (e) Block5-3 convolution layers of VGG19 network.

Figure 6

(a): The average ROC curves of the 20-times repeated five-fold cross-validation experiments of several classifiers trained by different features. (b): The histogram of AUC, ACC, TPR, and TNR, and the precision of the different classifiers. (c): Boxplot of the average AUC values of the classifiers constructed with the different features. A represents the classifier trained by clinical parameters; B represents the classifier trained by domain-specific features; C represents the classifier trained by clinical parameters and domain-specific features; D represents the classifier trained by abstract features; E represents the classifier trained by abstract features and clinical parameters; and F represents the classifier trained by abstract features, domain-specific features, and clinical parameters.