MRI-based Habitat Radiomics for Differentiating T1-Stage Nasopharyngeal Carcinoma From Benign Hyperplasia: A Dual-Center Study.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, The First Affiliated Hospital, Hengyang Medical School, University of South China, Hengyang, Hunan, China (H.Z.).
- Department of Radiology, Hainan Affiliated Hospital of Hainan Medical University (Hainan General Hospital), Haikou 570311, China (L.J.Z., T.C.L., H.T.L., M.Y.D., Y.H.G., W.Y.H.).
- Department of Radiology, Yueyang Central Hospital, Yueyang, Hunan, China (L.H.W.).
- Department of Radiology, Hainan Affiliated Hospital of Hainan Medical University (Hainan General Hospital), Haikou 570311, China (L.J.Z., T.C.L., H.T.L., M.Y.D., Y.H.G., W.Y.H.). Electronic address: [email protected].
Abstract
Differentiating T1-stage nasopharyngeal carcinoma (NPC) from benign hyperplasia (BH) on MRI remains challenging. This study aimed to develop and externally validate a habitat-based MRI radiomics framework for this diagnostic task and to evaluate its clinical utility through a multi-observer study. This retrospective dual-center study enrolled 194 patients with histopathologically confirmed T1-stage NPC or BH (internal cohort: 159; external cohort: 35). Following nnU-Net-assisted segmentation on non-contrast T1- and T2-weighted images, intratumoral habitats were generated via K-means clustering. Three support vector machine classifiers were trained and compared: whole-tumor radiomics (WTR), habitat radiomics (HR), and fusion radiomics (FR). Model interpretability was assessed using SHAP and LIME. Four radiologists with varying experience levels evaluated 82 test cases in a two-phase crossover observer study with and without an AI-assisted habitat report. FR achieved the highest AUC of 0.939 in the training set, 0.862 in the internal test set, and 0.863 in the external test set. HR consistently outperformed WTR despite relying on fewer features, with AUCs of 0.851 versus 0.840 in the internal test set and 0.843 versus 0.810 in the external test set. With AI assistance, diagnostic accuracy improved in all four radiologists from a range of 58.5-74.4% to 74.4-84.1%, net reclassification improvement reached significance in three of four radiologists, and Fleiss' κ improved from 0.083 to 0.329. Habitat-based MRI radiomics provided incremental diagnostic value for differentiating T1-stage NPC from BH using non-contrast sequences, and the AI-assisted report improved both diagnostic accuracy and inter-observer agreement.