Abstract 662P
Background
Precise recurrence detection facilitates timely intervention and prolongs survival in survivors with cured nasopharyngeal carcinoma (NPC). We presented an aide simultaneously reconciles dynamic risk probabilities profile and precise recurrence detection utilizing longitudinal magnetic resonance (MR) scans in NPC.
Methods
In this multicentre observational study, the Artificial Intelligence in detection of Recurrent Nasopharyngeal carcinoma (RAIN) was constructed by recurrent neural network on the basis of time-series head and neck MR scans from a retrospective development cohort from the Sun Yat-sen University Cancer Center (SYSUCC), Guangzhou, China. The accuracy of RAIN was validated in one internal (SYSUCC) and two external cohorts (set 1: the Cancer Hospital of The University of Chinese Academy of Sciences, Hangzhou, China; set 2: the Affiliated Cancer Hospital of Guangzhou Medical University, Guangzhou, China), and a prospective set (SYSUCC). A reader study between RAIN and three oncologists from the Fujian Provincial Cancer Hospital (Fujian, China) was conducted in the prospective cohort. Performances were evaluated using area under the curve (AUC).
Results
Between May 28, 2008, and Nov 24, 2020, 1183 patients with 6474 scans were retrospectively recruited in training cohort, 147 with 709 scans in tuning cohort, 147 with 780 scans in internal validation cohort, 403 with 3043 scans and 96 with 519 scans in external validation cohort 1,2; 248 eligible patients with 1780 scans were enrolled in the prospective cohort between Jan 5 and Dec 1, 2021. RAIN presents sound performance in internal validation cohort (AUC 0·805 [95% CI 0·780–0·830]), and external validation cohorts (1: 0·794 [0·746–0·842]; 2: 0·772 [0·734–0·810]). In the prospective validation cohort, RAIN achieved sensitivity similar to that of the expert oncologist (0·679 [95% CI 0·661–0·694] vs 0·712 [0·698–0·726]; p=0·529) and superior to competent (0·610 [0·589–0·628]; p<0·0001) and trainee (0·424 [0·410–0·440]; p<0·0001) oncologists.
Conclusions
RAIN was able to identify scan-detected recurrences with robustness and could therefore be implemented in clinical practice to optimize recurrence surveillance in NPC.
Clinical trial identification
Chinese Clinical Trial Registry ChiCTR2200056595).
Editorial acknowledgement
No.
Legal entity responsible for the study
The authors.
Funding
This study was supported by the National Natural Science Foundation of China (grant numbers 81872375, 82172862, and 82172863), and the Natural Science Foundation of Guangdong Province, China (grant number 2021A1515010118).
Disclosure
All authors have declared no conflicts of interest.