IMPACT OF NON-EXPERT LABELED DATASETS IN THE PERFORMANCE OF U-NET IN BIOMEDICAL IMAGE SEGMENTATION

Abstract

Image segmentation is introduced as partitioning an image into meaningful and disjointed regions offering a simplified representation of the image. It is a frontline domain of computer vision and one of the earliest problem statements considered by researchers. Despite the considerable number of available research, image segmentation remains a challenging endeavor in computer vision due to its significant technical challenges. The complex nature of this operation has made its implementations dependent on the quality and quantity of labeled data. The imperfection of the dataset especially in biomedical imaging would lead to the misinterpretation of such images during diagnosis. The purpose of this thesis is to make evident the deterioration of the performance of U-Net in segmenting biomedical images while using non-expert labeled datasets. Along with it, is to observe the behavior of U-Net while making certain adjustments in the datasets used and the implementation provided. Tested in three datasets, U-Net architecture behaves differently on datasets with different levels of label noise. Results from the conducted experiments have been examined from both qualitative and quantitative perspective. Nonetheless, it is worth mentioning that there exist alterations that can be applied to the dataset images prior to training phase that would contribute to a substantial improvement. However, such improvements are not sufficing, upholding so the fact that only experts’ annotations would result always in satisfactory and promising results. In addition, this thesis gives the reader a comprehensive view of the elevations of deep learning-based techniques in computer vision and in more details in medical image segmentation.