Development of an Automated Endoscopic Image Segmentation Technique for Accurate Localization of Gastrointestinal Lesions

Abstract

Automatic detection of gastrointestinal (GI) lesions makes endoscopic diagnosis more accurate, but it is difficult because the lesions can look different. A deep learning (DL) method for specific classification makes things more accurate, but for clinical use, a strong model is needed to make sure the location is correct. This study comes up with a better way to divide things up and solves these problems. The goal is to make a custom partition frame with a Residual Swin Transformer Fusion Network (RN-RSTFN) that has been fine-tuned with a Refined Nutcracker. Gaussian Filtering (GF) was used to reduce noise in the images and make them clearer. Then, Z-Score was used to standardize the distribution of pixel intensity. Function Extraction using a histogram of oriented gradients (HOG) helps to get the wound pattern needed for better partition performance. Give the right location for GI ulcers. Endoscopic image datasets that are available to the public from depot and medical institutions include pictures of different GI lesions. Gaussian Filtering (GF) was used to reduce noise in the images and make them clearer. Then, Z-Score was used to standardize the distribution of pixel intensity. The RN-RSTFN that was suggested combines the hierarchical representation and residual learning of the Swin Transformer to improve the boundaries of lesions. When measured using the Dice score and mIOU, Precision, and Recall measures, the RN-RSTFN model showed a lot of improvement in finding GI lesions. The scores were 0.9458, 0.9383, 0.9634, and 0.9596, respectively.