Improving Gastric Lesion Detection by using Specular Highlights Removal Algorithm and Deep Learning Approach

Gastric lesions cause many types of cancer with high mortality rates. Therefore, detecting gastric lesions early is necessary to prevent the risk of cancer. A conventional method used to detect these lesions is an endoscopy procedure. However, observation on the endoscopy images can be a challenge for endoscopists due to their irregular shapes and sizes, which can make them appear similar to the surrounding tissue, thus, increasing the likelihood of missing them during the procedure. Using deep learning models for automated lesion detection can help to reduce the rate of missed lesions by endoscopists. Recently, the most effective deep learning model for lesion detection is the You Only Look Once version 8 (YOLOv8) model. However, one of the challenging issues of deep learning models is to handle the specular highlight areas on the endoscopy images that make noise to the object detection model. To handle this, we propose the preprocessing image method named Specular Highlights Removal (SHR) algorithm to eliminate the specular highlights areas to improve the quality of endoscopy images. As a result, our proposed solution enhances the accuracy of the deep learning model for the lesion detection problem. This is proved by the extensive experiments on two collected datasets, i.e., Negative Helicobacter-Pylori(HP) and Positive HP datasets.