Deep Learning MRI-based Radiomics:Prediction Of Distant Metastasis Risk In Nasopharyngeal Carcinoma

ObjectiveTo evaluate the value of a deep learning model for metastasis risk prediction based on MRI images of positive lymph nodes in the neck before treatment in patients with nasopharyngeal carcinoma(NPC).MethodThis study retrospectively included patients with nasopharyngeal carcinoma who were newly diagnosed without metastasis and were node-positive,and all patients had pre-treatment nasopharyngeal + neck MRI(T1W and T2W)images,and a physician using 3D-Slicer software delineated the region of interest(metastatic lymph nodes)in two MRI sequences.In this study,a 3D-Unet deep learning network was used to build a transfer prediction deep learning(DL)model.The main observation endpoint was Distant Metastasis-Free Survival(DMFS),the Kaplan-Meier curve was plotted to calculate the survival rate,and Harrell’s Concordance index(C-index)was used to evaluate the performance of the deep learning model.OutcomeA total of 678 patients with newly diagnosed NPC without metastasis(training cohort: N=291;validation cohort: N=387)were included in the study,and the metastasis risk prediction model constructed by positive lymph node MRI in the neck before treatment was included,and the transfer risk score of T1 W combined with T2 W was calculated and the transfer risk stratification was performed.In the training cohort,patients in the high-metastasis-risk group had significantly inferior 5-year DMFS to those in the low-metastasis group(38.5% vs 96.2%,P<0.001),and was also observed in the validation cohort(74.6% vs 87.7%,P<0.001).In both the training pair column and the validation cohort,multivariate analysis showed independent prognostic factors with high-risk scores for DMFS and OS.In terms of transfer risk assessment,the transfer risk score performed significantly better than N classification and EBV-DNA(training cohort: 0.832 vs 0.619 vs 0.627,P< 0.001;validation cohort:0.678 vs 0.577 vs 0.521,P<0.001).In the high-metastasis risk group,patients with radiotherapy alone had a significantly higher risk of metastasis than those in the chemotherapy ≦ 3 and > 3 cycles(5-year DMFS: 30.2% vs 54.5% vs 66.0%,P=0.047);In the low-risk group,there was no significant difference between radiotherapy alone and chemotherapy ≦ 3 cycles and 5-year DMFS > cycles(90.1versus 94.1 versus 94.4 percent,P = 0.42).ConclusionThe prognostic model constructed based on deep learning MRI images of positive lymph nodes in the neck can effectively predict the risk of nasopharyngeal cancer metastasis,and can be used as a decision-making tool for individualized treatment of NPC patients.