A DEEP LEARNING APPROACH FOR SEGMENTATION OF NOISY MEDICAL IMAGES USING SCD-U-NET

Abstract

Image segmentation plays a vital role in clinical decision-making within the medical field, significantly contributing to the sustainability of healthcare. As a result, the study of biomedical image segmentation has emerged as a key focus area in computer vision research. Analysis of medical images frequently necessitates the segmentation of objects, yet the availability of training data for this particular activity is typically limited and challenging to acquire. So, training for object segmentation in medical image analysis is hard. Further, the noise in the acquired data might not always give optimum segmentation results. Hence, training the data becomes a two-fold challenge. Therefore, we present SCD-U-Net based Seg-mentation, a novel method that requires only a small number of annotated ground truth segmentation to train it from end to end. Our method is based on the self-supervised denoising that works even in the absence of a clean pair of images.