Assessing the Performance of U-Net in 3D Medical Image Segmentation

Despite the increasing speed of development in 3D medical image segmentation based on deep learning techniques, it is important to consider the requirements for performing this process due to the large number of information contained in the data. So far, researchers have proposed several methods to avoid excessive demand placed on computational resources. The objective of this work is to address this issue by leveraging 2D based Convolutional Neural Network (CNN) architectures to segment 3D medical images in order avoid resource overload. To achieve this, it has been divided the work into three phases. The first one consists of splitting each 3D image alongside depth axis. Next, it has been trained the extracted slices through U-Net; one of the most common CNN architectures in the biomedical image segmentation field. To end up with reconstructing the predicted slices back to volumetric images. In the MSD-Spleen dataset, our best F1-Score, IoU, and Accuracy on the validation set were 0.840 and 0.89 respectively. Moreover, this method has shown efficiency to train volumetric images using limited resources.

I. Introduction

Image segmentation in medical imaging refers to the process of dividing images into regions or parts corresponding to specific organs or tissues; it is a crucial task in various medical applications. Medical images, particularly three-dimensional ones, are complex and information-rich, making their segmentation challenging [1]. Thanks to recent advancements in artificial intelligence, precisely deep learning [2], which has been shown effective results compared to traditional learning methods. There has been needed for robust techniques to reduce time and guide professionals in various healthcare applications [3].