Radiomic Analysis for Pretreatment Prediction of Recurrence after Radiotherapy in Locally Advanced Cervical Cancer.

Purpose/objective(s) In earlier stage of locally advanced unresectable cervical squamous cell carcinoma, the treatment outcomes of surgery and definitive radiotherapy are comparable. However, the past study reported that one third of patients would experience recurrence. Predicting the treatment response and the long-term treatment outcome presents a major challenge for developing a more precise personalized care. Radiomics is a non-invasive and an emerging low-cost method to predict prognosis. In the current study, the recurrence of cervical cancer patients treated with radiotherapy was predicted using the radiomics features from the extended and shrink-uterus regions that extracted from pre-treatment T1- and T2-weighted MRI images. Materials/methods Data of 90 tumor were split into two sets: 67 tumors for the training of models and 23 tumors for model testing. The treatment outcome was classified into two groups. The first group (group I) was patients with a recurrence. The second group (group II) was patients without a recurrence. A total number of 5022 radiomics features per a patient image were extracted from normalized and unnormalized T1- and T2-weighted MRI images. The set of candidate predictors were selected with the least absolute shrinkage and selection operator (LASSO) logistic regression and build predictive models with neural network classifiers were used. The precision, accuracy, and sensitivity by generating confusion matrices and the areas under the receiver operating characteristic curve (AUC) for each model were evaluated. Results By the LASSO analysis of the training data, we found 25 radiomics features from unnormalized T1-weighted (unT1w) MRI image and 4 radiomics features from unT2w MRI image for the classification. On the other hand, 11 radiomics features from normalized T1-weighted (nT1w) MRI image and 27 radiomics features from nT2w MRI image for the classification. The accuracy, specificity, sensitivity, and AUC of the prediction model for the dataset were 86.4%, 74.9%, 81.8%, and 0.89 with unT1w MRI image, 89.4%, 38.1%, 72.2%, and 0.69 with unT2w MRI image, 93.1, 81.6%, 88.7%, and 0.94 with the combination of unT1w and unT2w MRI images. On the other hand, the accuracy, specificity, sensitivity, and AUC of the prediction model for the dataset were 83.0%, 72.2%, 89.1%, and 0.90 with nT1w MRI image, 93.3%, 90.8%, 94.5%, and 0.96 with nT2w MRI image, 96.3, 92.8%, 98.9%, and 1.00 with the combination of nT1w and nT2w MRI images. Conclusion The radiomics features of the difference of the pixel number at center and peripheral region and the variation of the distribution were important factors in predicting the recurrence of the cervix cancer. The normalized method improved the accuracy of the prediction of the recurrence for cervix cancer.