Deep Learning vs. Handcrafted Radiomics to Predict Chemoradiotherapy Response for Locally Advanced Cervical Cancer.

Abstract

To predict CRT response in locally advanced cervical cancer (LACC) with handcrafted radiomics (HCR) and deep learning radiomics (DLR) using pretreatment MRI. Furthermore, we investigate whether the incorporation of clinical factors improves prediction performance. Two hundred and fifty-two patients with LACC are enrolled. All patients are treated with external beam radiotherapy, followed by high-dose-rate intracavitary brachytherapy with concurrent cisplatin. The patients are randomly divided into two independent groups for the training (167 patients) and test datasets (85 patients). Contrast-enhanced T1- and T2-weighted MR scans are obtained. Patients in the training and test sets have similar characteristics in terms of age, tumor size, FIGO stage, HPV infection status, or CRT response. For HCR analysis, 1890 imaging features are extracted and a support vector machine classifier with a five-fold cross-validation is trained using training dataset to predict CRT response and validated using test dataset. For DLR analysis, a 3-dimensional convolutional neural network was trained and validated using test dataset. A comparison of the DLR and HCR models reveals that the DLR model exhibits better prediction performance than the HCR model for the test dataset (AUC = 0.721 vs. 0.597, p = 0.097). The incorporation of clinical factors could improve performance in both DLR and HCR models. The DLR models outperform the HCR models in predicting CRT responses in patients with LACC. Combining clinical factors and MRI may improve the prediction performance in both HCR and DLR analyses.