面向深度学习的胰腺医学图像分割方法研究进展

Research Progress on Pancreatic Medical Image Segmentation Methods Based on Deep Learning

精确的胰腺医学图像分割在胰腺疾病的术前诊断,手术计划以及术后评估中均具有重要意义.相较于肝脏和肺,胰腺在不同个体上的形态体积具有高度的可变性,因此准确自动的分割胰腺是一项具有挑战性的任务.近年来,深度学习为胰腺分割提供了一种高效率、高精度的解决方案,较之于传统方法有明显的性能提升.本文研究回顾了近五年基于深度学习的胰腺分割领域的相关文献,梳理了常用的胰腺分割数据集,并对胰腺的深度学习分割方法进行了较为详尽的分类与总结.重点介绍了每类分割方案的原理、基本思想、网络架构,评述了方案的优缺点,并在统一评价指标上进行分割性能的比较.最后本文提出了现有的基于深度学习的胰腺分割方法存在的问题,并对未来的研究趋势进行了展望.

In the preoperative diagnosis,surgical planning,and postoperative assessment of pancreatic disorders,accurate segmentation of pancreatic medical pictures is critical.The medical imaging features of the pancreas are characterized by blurred organ-tissue boundaries,uneven pixels within the tissue,and high variability in the morphological volume of organs on different individuals,in contrast to abdominal organs such as the liver and lung,which have obvious organ boundaries and rigid geometric features.As a result,compared to the liver and lung,the pancreatic segmentation accuracy is poor,making it difficult to properly and automatically segment the pancreas of various individuals.Deep learning approaches have proved their remarkable feature learning capabilities in recent years,and they clearly outperform classic segmentation methods in terms of performance.As a result,researchers use deep learning to segment pancreatic medical images,giving an efficient and high-precision solution for the area of pancreatic segmentation.In recent years,it has become the standard approach for pancreatic medical image segmentation.This work analyzes the significant literature in the field of pancreatic segmentation using deep learning during the last five years,classifies and describes the deep learning pancreas segmentation techniques in depth,and sorts out the most often used pancreas segmentation datasets.Each type of segmentation scheme′s premise,fundamental idea,and network design are basically explained,the benefits and drawbacks of different schemes are discussed,and the segmentation performance of each segmentation scheme is compared on the unified evaluation index.Finally,this work addresses the shortcomings of existing deep learning-based pancreatic segmentation algorithms,as well as highlights and forecasts the future research direction in pancreas segmentation.