Development and Validation of CT-Based Clinical-Radiomics Nomogram for Early Stage Extranodal Nasal-Type NK/T Cell Lymphoma: A Multicenter Study

Abstract

Most patients with extranodal nasal-type NK/T cell lymphoma (ENKTCL) had a localized disease with extensive primary tumor invasion at diagnosis (70-90%). Several clinical risk indexes, such as nomogram-revised risk index (NRI), international prognostic index (IPI), Korean Prognostic Index (KPI) and prognostic index of natural killer lymphoma (PINK), were used for ENKTCL patient stratification and providing information in clinical decision-making. However, they had low predictive power for early-stage patients with ENKTCL. This is the first study to construct a model with more predictive power through CT-based radiomics signature combined with traditional clinical risk indexes for overall survival (OS) of patients with early-stage ENKTCL. A total of 196 early stage ENKTCL patients were randomly assigned into the training (n = 147) and interval validation set (n = 49) in a 3:1 ratio. And 83 and 19 early stage ENKTCL patients from other two centers were used for external validation set (n = 62). All patients received radiotherapy after 2-3 cycles of chemotherapy. 1316 CT radiomic features before radiotherapy were extracted and selected to construct the radiomics signature (RS). A CT-based nomogram was established by integrating clinical indexes and radiomics signature in training set and was tested in two validation sets. With a median follow-up period of 59.9 months, 48 patients (24.1%) died. Compared with other prognostic index, NRI had better power to predict 5-year OS in the training cohort. The radiomics signature constructed by 11 selected radiomic features showed better prognostic performance than NRI for predicting 5-year OS in training set (C-index: 0.75 vs. 0.66), internal validation set (C-index: 0.71 vs. 0.62) and external validation set (C-index: 0.68 vs. 0.60). Patients were stratified into high- and low-risk groups by median radiomic signature. Patients in high-risk group had worse 5-year OS than patients in low-risk group (training set: 92% vs. 65%, P<0.001; internal validation set: 88% vs. 59%, P<0.05; external validation set 90% vs. 60%, P<0.05). The nomogram established by integrating radiomics signature with NRI showed optimal prognostic performance with C-index of 0.77 in training, 0.73 in internal and 0.71 in external validation set. Calibration curves showed good agreement. The clinical-radiomics nomogram integrating CT-based radiomics signature combined with traditional clinical risk index provided an excellent prognostic tool for OS, which could be helpful for personalized risk stratification and treatment in early stage ENKTCL patients.