Deep learning for the precise detection of recurrence in nasopharyngeal carcinoma from time-series medical imaging

Abstract Precise detection of recurrence in patients with treated nasopharyngeal carcinoma (NPC) facilitates timely intervention and prolongs survival. However, there is no compelling tool realizing real-time precise recurrence detection as scale hitherto. Here we present a deep learning-based sequential scan model called RAIN, harnessing 10,212 time-series follow-up head and neck magnetic resonance (MR) scans of 1,808 patients with treated NPC in a multicenter observational study (Blinded ID). The RAIN yields larger area under the receiver operating curve (AUC) values than single scan model (internal: 0.916 vs 0.855, p = 0.004; external: 0.900 vs 0.709, p < 0.001). The reader study showed RAIN has superiority in timely detection of recurrence than readers. These findings suggested that RAIN could detect recurrence on MR scans with high precision and therefore be implemented in clinical practice to optimize recurrence surveillance in treated NPC.