• Shuangshuang Wu, Haojiang Li, Annan Dong, Li Tian, Guangying Ruan, Lizhi Liu, and Yuanzhi Shao.
• School of Physics, State Key Laboratory of Optoelectronic Materials and Technologies, Sun Yat-sen University, Guangzhou, PR China.
• J Magn Reson Imaging. 2021 Sep 1; 54 (3): 854-865.

BackgroundAccurately predicting the risk of death, recurrence, and metastasis of patients with nasopharyngeal carcinoma (NPC) is potentially important for personalized diagnosis and treatment. Survival outcomes of patients vary greatly in distinct stages of NPC. Prognostic models of stratified patients may aid in prognostication.PurposeTo explore the prognostic performance of MRI-based radiomics signatures in stratified patients with NPC.Study TypeRetrospective.PopulationSeven hundred and seventy-eight patients with NPC (T1-2 stage: 298, T3-4 stage: 480; training cohort: 525, validation cohort: 253).Field Strength/SequenceFast-spin echo (FSE) axial T1-weighted images, FSE axial T2-weighted images, contrast-enhanced FSE axial T1-weighted images at 1.5 T or 3.0 T.AssessmentRadiomics signatures, clinical nomograms, and radiomics nomograms combining the radiomic score (Radscore) and clinical factors for predicting progression-free survival (PFS) were constructed on T1-2 stage patient cohort (A), T3-4 stage patient cohort (B), and the entire dataset (C).Statistical TestsLeast absolute shrinkage and selection operator (LASSO) method was applied for radiomics modeling. Harrell's concordance indices (C-index) were employed to evaluate the predictive power of each model.ResultsAmong 4,410 MRI-extracted features, we selected 16, 16, and 14 radiomics features most relevant to PFS for Models A, B, and C, respectively. Only 0, 1, and 4 features were found overlapped between models A/B, A/C, and B/C, respectively. Radiomics signatures constructed on T1-2 stage and T3-4 stage patients yielded C-indices of 0.820 (95% confidence interval [CI]: 0.763-0.877) and 0.726 (0.687-0.765), respectively, which were larger than those on the entire validation cohort (0.675 [0.637-0.713]). Radiomics nomograms combining Radscore and clinical factors achieved significantly better performance than clinical nomograms (P < 0.05 for all).Data ConclusionThe selected radiomics features and prognostic performance of radiomics signatures differed per the type of NPC patients incorporated into the models. Radiomics models based on pre-stratified tumor stages had better prognostic performance than those on unstratified dataset.Level Of Evidence4 Technical Efficacy Stage: 5.© 2021 International Society for Magnetic Resonance in Medicine.

15 keywords...

hide…