Machine learning-based radiomics for predicting response to concurrent chemoradiotherapy in locally advanced cervical cancer: An exploratory study

Objective This study aimed to investigate the value of machine learning models based on pre-treatment radiomics for predicting response to concurrent chemoradiotherapy (CCRT) in locally advanced cervical cancer (LACC) patients. Response was defined as a tumor regression rate reaching 90% before brachytherapy. Methods This retrospective exploratory study analyzed 50 patients with locally advanced cervical cancer who underwent concurrent chemoradiotherapy. Radiomic features were extracted from primary lesions using pre-treatment contrast-enhanced T1-and T2-weighted imaging sequences. Tumor contours were delineated on T2-weighted magnetic resonance imaging images using the Monaco treatment planning system. Patients were classified according to whether the tumor regression rate (TRR) reached 90% before brachytherapy after concurrent chemoradiotherapy-high (TRR ≥ 90%)/low (TRR < 90%) regression groups. Machine learning models were constructed, and their performance was evaluated using the area under the curve (AUC) and confusion matrices. Results We constructed the model based on contrast-enhanced T1-weighted imaging (T1WI) images. The results showed that among the models based on pre-treatment contrast-enhanced T1WI images, the random forest (RF) model achieved the highest AUC of 0.760, outperforming the Tree (0.637), support vector machine (0.713), and Logit (0.624) models. The confusion matrix for the RF model showed an accuracy of 71.3%, a precision of 70.4%, and a recall of 76.7%. Furthermore, we evaluated the performance of the model on T2WI. Among the models based on pre-treatment T2WI images, the AUC of single models (SVM, 0.845; RF, 0.865; gradient boosting, 0.863; LR, 0.813). The advanced fusion model demonstrated the best overall performance with a test set AUC of 0.862 and cross-validation AUC of 0.863. Conclusion Machine learning based on pre-treatment radiomics may help predict whether the TRR will reach ≥90% after CCRT in patients with LACC. It shows preliminary potential as an auxiliary tool that could assist in informing treatment decisions and might help avoid unnecessary dose escalations.