One of the major difficulties in medical image segmentation is the high variability of these images, which is caused by their origin (multi-centre), the acquisition protocols (multi-parametric), as well as the variability of human anatomy, the severity of the illness, the effect of age and gender, among others. The problem addressed in this work is the automatic semantic segmentation of lumbar spine Magnetic Resonance images using convolutional neural networks. The purpose is to assign a class label to each pixel of an image. Classes were defined by radiologists and correspond to different structural elements like vertebrae, intervertebral discs, nerves, blood vessels, and other tissues. The proposed network topologies are variants of the U-Net architecture. Several complementary blocks were used to define the variants: Three types of convolutional blocks, spatial attention models, deep supervision and multilevel feature extractor. This document describes the topologies and analyses the results of the neural network designs that obtained the most accurate segmentations. Several of the proposed designs outperform the standard U-Net used as baseline, especially when used in ensembles where the output of multiple neural networks is combined according to different strategies.
翻译:医疗图象分解的主要困难之一是这些图象因来源(多中心)、获取协议(多参数)以及人类解剖的变异性、疾病的严重程度、年龄和性别的影响等原因造成的高变异性,在医学图象分解过程中遇到的一个重大困难是,在这项工作中涉及的问题是,使用卷发神经网络对卢巴脊椎磁共振成像的图象自动进行语解分解,目的是为图像的每像像像标定一个等级标签,由放射学家确定等级,并对应诸如脊椎、间垂直盘、神经、血管、血管和其他组织等不同结构要素。拟议的网络表象是U-Net结构的变异。使用几个互补区块来界定变异性:三种类型的变动区块、空间关注模型、深层监督和多级特征提取器。本文件描述了最精确的神经网络设计的表象学和分析结果。若干拟议的设计超越了标准的U-Net网络的模型,在使用不同的输出基线时,特别是用于不同的数字网络。