Medical Image Segmentation using modified U-Ne

Abstract

  Image analysis ranging from lesions to tissues, organs, or entire bodies is central in solving medical issues by pulling out essential data from images acquired through medical imaging facilities so as to enhance odds of an accurate diagnosis. For our model described in this project, we propose a new network structure extending the U-Net model called the Image Contrast-Based U-Net Segmentation model that reframed to diagnose skin diseases by identifying affected areas of the skin. Using red rectangles the model emphasizes the affected areas which is extremely helpful for physicians when diagnoses different types of skin diseases. Through experiments using the ISIC and PH2 datasets, the U-Net model is trained to determine the accuracy above 98% of affected pixels. Also, the specificity of the provided system is the ability to upload test images, as well as to observe the initial and segmented images with marked useful areas and damages. The performance comparison of the two models reveals similar efficiency when working with the two sets of data. The proposed system provides the means for easy identification of skin diseases through simple efficient tools for segmentation of images and can be generalized for improved implementation of medical image segmentation for diagnosis in clinical environments.