Deep learning for real-time detection of nasopharyngeal carcinoma during nasopharyngeal endoscopy

Zicheng He; Kai Zhang; Nan Zhao; Yongquan Wang; Weijian Hou; Qinxiang Meng; Chunwei Li; Junzhou Chen; Jian Li

doi:10.1016/j.isci.2023.107463

Deep learning for real-time detection of nasopharyngeal carcinoma during nasopharyngeal endoscopy

iScience. 2023 Jul 24;26(10):107463. doi: 10.1016/j.isci.2023.107463. eCollection 2023 Oct 20.

Authors

Zicheng He^{1

2}, Kai Zhang³, Nan Zhao³, Yongquan Wang¹, Weijian Hou⁴, Qinxiang Meng⁵, Chunwei Li¹, Junzhou Chen³, Jian Li^{1

2}

Affiliations

¹ Department of Otolaryngology, the First Affiliated Hospital of Sun Yat-sen University, Guangzhou Key Laboratory of Otorhinolaryngology, Otorhinolaryngology Institute of Sun Yat-sen University, Guangzhou 510080, P.R.China.
² Guangxi Hospital Division of The First Affiliated Hospital, Sun Yat-sen University, Nanning 530000, P.R.China.
³ School of Intelligent Systems Engineering, Shenzhen Campus of Sun Yat-sen University, Shenzhen 518107, P.R.China.
⁴ Kiang Wu Hospital, Macau 999078, P.R.China.
⁵ Guangzhou First People's Hospital, Guangzhou 510180, P.R.China.

Abstract

Nasopharyngeal carcinoma (NPC) is known for high curability during early stage of the disease, and early diagnosis relies on nasopharyngeal endoscopy and subsequent pathological biopsy. To enhance the early diagnosis rate by aiding physicians in the real-time identification of NPC and directing biopsy site selection during endoscopy, we assembled a dataset comprising 2,429 nasopharyngeal endoscopy video frames from 690 patients across three medical centers. With these data, we developed a deep learning-based NPC detection model using the you only look once (YOLO) network. Our model demonstrated high performance, with precision, recall, mean average precision, and F1-score values of 0.977, 0.943, 0.977, and 0.960, respectively, for internal test set and 0.825, 0.743, 0.814, and 0.780 for external test set at 0.5 intersection over union. Remarkably, our model demonstrated a high inference speed (52.9 FPS), surpassing the average frame rate (25.0 FPS) of endoscopy videos, thus making real-time detection in endoscopy feasible.

Keywords: Cancer systems biology; Machine learning.