MRI-based deep learning and radiomics for predicting the efficacy of PD-1 inhibitor combined with induction chemotherapy in advanced nasopharyngeal carcinoma: A prospective cohort study

Abstract

Background

An increasing number of nasopharyngeal carcinoma (NPC) patients benefit from immunotherapy with chemotherapy as an induction treatment. Currently, there isn't a reliable method to assess the efficacy of this regimen, which hinders informed decision-making for follow-up care.

Aim

To establish and evaluate a model for predicting the efficacy of programmed death-1 (PD-1) inhibitor combined with GP (gemcitabine and cisplatin) induction chemotherapy based on deep learning features (DLFs) and radiomic features.

Methods

Ninety-nine patients diagnosed with advanced NPC were enrolled and randomly divided into training set and test set in a 7:3 ratio. From MRI scans, DLFs and conventional radiomic characteristics were recovered. The random forest algorithm was employed to identify the most valuable features. A prediction model was then created using these radiomic characteristics and DLFs to determine the effectiveness of PD-1 inhibitor combined with GP chemotherapy. The model's performance was assessed using Receiver Operating Characteristic (ROC) curve analysis, area under the curve (AUC), accuracy (ACC), and negative predictive value (NPV).

Results

Twenty-one prediction models were constructed. The Tf_Radiomics+Resnet101 model, which combines radiomic features and DLFs, demonstrated the best performance. The model's AUC, ACC, and NPV values in the training and test sets were 0.936 (95%CI: 0.827–1.0), 0.9, and 0.923, respectively.

Conclusion

The Tf_Radiomics+Resnet101 model, based on MRI and Resnet101 deep learning, shows a high ability to predict the clinically complete response (cCR) efficacy of PD-1 inhibitor combined with GP in advanced NPC. This model can significantly enhance the treatment management of patients with advanced NPC.