Evaluating Hyperspectral Techniques Using Objective Metrics Research on Analog Narrowband Image

Abstract

Abstract The evident signs of esophageal cancer (EC) typically do not become noticeable until the middle or late stages. The survival rate of EC is reduced to less than 20% if it is detected in the latter stages. This paper compares the performance of white light image (WLI), narrowband imaging (NBI), cycle-consistent adversarial network (CycleGAN) simulated narrowband image (CNBI), and hyperspectral imaging (HSI) simulated narrowband image (HNBI) to detect EC in its early stages. A total of 1000 EC images (500 WLI images and 500 NBI images) were used as dataset in collaboration with Kaohsiung Armed Forces General Hospital. The CycleGAN model was used to produce CNBI. An HSI imaging algorithm was also developed to produce HNBI images. The effectiveness of these four types of images in detecting EC at its early stages was evaluated based on three indicators, namely, CIEDE2000, entropy, and structural similarity index measure (SSIM). Results of CIEDE2000, entropy, and SSIM analysis suggest using CycleGAN to generate CNBI and HNBI images is superior in detecting EC compared with normal WLI and NBI.