AUTHOR=Pan Mulan , Lu Lu , Mu Xingyu , Jin Guanqiao 

TITLE=Prediction of induction chemotherapy efficacy in patients with locally advanced nasopharyngeal carcinoma using habitat subregions derived from multi-modal MRI radiomics

JOURNAL=Frontiers in Oncology

VOLUME=Volume 15 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2025.1539574

DOI=10.3389/fonc.2025.1539574

ISSN=2234-943X

ABSTRACT=ObjectiveThis study aims to predict the early efficacy of induction chemotherapy (ICT) in patients with locally advanced nasopharyngeal carcinoma (LA-NPC) through habitat subregion analysis and multimodal MRI radiomics techniques.MethodsThe study employed a retrospective design and included LA-NPC patients who received ICT treatment between 2015 and 2019. The K-means clustering algorithm was utilized to segment the tumor into five distinct habitat subregions based on imaging features. A total of 2,153 radiomic features, including geometric shape, intensity, and texture features, were extracted. Feature selection was conducted using the maximum relevance minimum redundancy (mRMR) method and the least absolute shrinkage and selection operator (LASSO) technique. Eleven machine learning algorithms were employed to develop radiomics models based on the CE-T1WI and T2-FS sequences, respectively. These models were evaluated using various predictive performance metrics, including area under the curve (AUC), sensitivity, and specificity. Model selection was based on comprehensive cross-validation performance and AUC values.ResultsThe study population comprised 76.63% males and 23.37% females, with a mean age of 42.60 ± 10.21 years. All patients had stage III to IVa nasopharyngeal carcinoma, and the majority (92.39%) had non-keratinizing squamous cell carcinoma. Habitat subregion analysis revealed that the volume features of a specific subregion (Subregion 2) were significantly associated with patient response to ICT (P = 0.032). The RF model built using radiomic features from Subregion 2 demonstrated the best performance on the CE-T1WI sequence, with an AUC of 0.921 in the training set and 0.819 in the testing set. On the T2-FS sequence, the Random Forest (RF) model also exhibited high diagnostic performance, with an AUC of 0.933 in the training set and 0.829 in the testing set. These results suggest that the RF model provides stable and reliable predictive performance across different MRI sequences.ConclusionHabitat subregion analysis using multimodal MRI radiomics offers an effective approach for the early identification of LA-NPC patients with poor responses to induction chemotherapy. This method holds promise for supporting clinical treatment decisions and achieving personalized medicine.