The problem
Endoscopic appearance varies with imaging modality. The study examines translating white-light images into another appearance while retaining polyp and inner-wall structure to supplement detector training.
The research idea
Unpaired MUNIT translates white-light endoscopy into images resembling narrow-band imaging (NBI), without requiring paired images. Generated images are added to SSD polyp-detector training.
What was evaluated
The publisher's public abstract reports detection evaluation using actual endoscopic polyp images. The study combines image generation with evaluation of a detector trained with generated samples.
Considering applications
The approach uses differences in imaging appearance for data augmentation. Preservation of lesion structure and detection performance on real images independent of generated training samples provide evaluation criteria.
Original research
Improvement of Polyp Detection using MUNIT for Image Generation
Iwahori, Yuji, Ooto, Tsubasa, Usami, Hiroyasu, Fukui, Shinji, Bhuyan, M. K., Wang, Aili, Ogasawara, Naotaka, Kasugai, Kunio
Procedia Computer Science, 2022 · KES 2022