This study utilized 248 white light endoscopic images and 258 narrow-band endoscopic images of early-stage esophageal cancer provided by Kaohsiung Medical University. After data processing and annotation, the images underwent augmentation and hyperspectral transformation. The dataset was then divided into training and testing sets and fed into the YOLOv5 model for the detection of esophageal cancer lesions. Four models were developed to predict the location of esophageal cancer lesions: separate models for white light and narrow-band images, as well as models for RGB and hyperspectral images. The results were evaluated based on the severity of the images, categorized as normal, dysplasia, and squamous cell carcinoma (SCC).According to the analysis of Dysplasia sensitivity, the sensitivity was 70.4% for white light esophageal cancer images, 79.5% for narrow-band esophageal cancer images, 84.3% for white light hyperspectral esophageal cancer images, and 84.5% for narrow-band hyperspectral esophageal cancer images. According to the analysis of SCC sensitivity, the sensitivity was 75.8% for white light esophageal cancer images, 81.3% for narrow-band esophageal cancer images, 86.6% for white light hyperspectral esophageal cancer images, and 86.9% for narrow-band hyperspectral esophageal cancer images. The results indicated that the model performed better in detecting SCC than Dysplasia, and narrow-band images had higher accuracy compared to white light images.Regarding mAP analysis, the mAP was 72.7% for white light esophageal cancer images, 86.8% for narrow-band esophageal cancer images, 78.6% for white light hyperspectral esophageal cancer images, and 88.6% for narrow-band hyperspectral esophageal cancer images. In conclusion, hyperspectral images demonstrated higher mAP values in both white light and narrow-band settings, indicating a significant improvement in the detection of esophageal cancer lesions compared to RGB images in this study.
|