Purpose: To establish an image-based M1 category subdivision system for personalized prognosis prediction and treatment planning in patients with metastatic nasopharyngeal carcinoma (NPC). Materials and Methods: A total of 1172 patients with metachronous metastasic NPC were retrospectively enrolled (the dataset is from Sun Yat-sen University Cancer Center for derivation, and the combined datasets are from Guangzhou Medical University Cancer Center and the Fifth Affiliated Hospital of Sun Yat-sen University for validation). The Ethics Committee of the three centers approved this study. A general subdivision system of the M1 category was established on the basis of the most influential metastatic features for overall survival (OS). The following multilevel subdivision system for precise subdivision of the M1 category was designed: M [number of locations]-Location [number of lesions], with B indicating bone, L indicating the lung, H indicating the liver, and N indicating a node. The correlation of the M1 subdivisions with OS was determined with Cox regression. The best treatment response was assessed with Response Evaluation Criteria in Solid Tumors 1.1 guidelines and modified Response Evaluation Criteria in Solid Tumors criteria. Results: Multivariate analysis in the derivation cohort showed that the number of metastatic lesions (multiple or single), the number of metastatic locations (multiple or single), liver involvement, and bone involvement were independent prognostic factors for OS. In general, subdividing the cohort by the number of metastatic lesions and the number of metastatic locations resulted in three subcategories of differential OS: M1a, a single lesion in a single organ or location; M1b, multiple lesions in a single organ or location; and M1c, metastases in multiple locations (for M1b vs M1a, hazard ratio [HR] = 2.28, 95% confidence interval [CI]: 1.71, 3.05; for M1c vs M1a, HR = 3.65, 95% CI: 2.75, 4.85); these subdivisions were externally validated. The multilevel subdivision system could be further used to discriminate among subgroups of differential OS under the M1b subcategory. Findings from analysis of multilevel subgroups suggested that patients with a single metastatic lesion (M1-B-1, M1L(1), M1-H-1, M1-N-1) or two lesions in the liver only (M1-H-2) had high rates of complete response (CR) or complete surgical resection (CSR) and 3-year OS after treatment (CR plus CSR rates >30%, and 3-year OS rates >50%); there were high 3-year OS rates (>50%) in patients with stage M1-B-2, M1-L-2, or M1-H-3 disease but relatively low rates of CR or CSR. Conclusion: Use of the multilevel M1 subdivision system in patients with NPC could facilitate more precise prognosis prediction and better identification of patients who will respond well to treatment than the conventional subdivision strategy. (C) RSNA, 2016