Deep Learning Improves Diagnostic Performance of Lymph Node Metastasis in Cervical Cancer by Using Magnetic Resonance Imaging

Background: Accurate identification of lymph node metastasis (LNM) preoperatively in patients with cervical cancer (CC) can avoid unnecessary surgical intervention and benefit treatment planning. In this study, we aimed to develop a deep learning (DL) model using magnetic resonance imaging (MRI) for non-invasive and preoperative LNM prediction in CC.Methods: This research retrospectively enrolled 479 patients with CC from three centres, and divided patients into a primary cohort (n= 338 from two centres) and an independent validation cohort (n= 141 from another centre). We proposed an end-to-end DL model to identify LNM in CC using MRI, compared the performance among three MRI sequences, and explored the effect of intratumoural and peritumoural regions. Moreover, we built a hybrid model to combine the DL model and clinical information, and assessed its prognostic value in predicting disease-free survival (DFS) of CC by Kaplan-Meier analysis.Findings: Among the three sequences, the DL model which used contrast-enhanced T1-weighted imaging and combined both intratumoural and peritumoural regions (defined as CET1WI tumour+ peri) showed the best performance (AUC= 0.844 in the validation cohort, P< 0.0001 between the positive and negative LNM patients). These results were further improved in the hybrid model that combined CET1WI tumour+ peri; and clinical lymph node status (AUC= 0.933 in the validation cohort). Moreover, the H-score from the hybrid model was significantly associated with DFS of CC.Interpretation: DL improves the diagnostic performance of LNM in patients with CC. The hybrid model can …