Wang Lichao, Merrifield Robert, Yang Guang-Zhong
The Hamlyn Centre for Robotic Surgery, Imperial College London, UK.
Med Image Comput Comput Assist Interv. 2011;14(Pt 3):627-34. doi: 10.1007/978-3-642-23626-6_77.
The ability to learn from user behavior during image segmentation to replicate the innate human ability to adapt shape delineation to contextually specific local information is an important area of study in image understanding. Current approaches to image segmentation usually incorporate specific designs, either relying on generic image features or specific prior knowledge, which usually prevent their application in different contextual settings. In this paper, a general segmentation framework based on reinforcement learning is proposed. It demonstrates how user-specific behavior can be assimilated in-situ for effective model adaptation and learning. It incorporates a two-layer reinforcement learning algorithm that constructs the model from accumulated experience during user interaction. As the algorithm learns 'pervasively' whilst the user performs manual segmentation, no additional steps are required for the training process, allowing the method to adapt and improve its accuracy as experience is acquired. Detailed validation of the method on in-vivo magnetic resonance (MR) data demonstrates the practical value of the technique in significantly reducing the level of user interaction required, whilst maintaining the overall segmentation accuracy.
在图像分割过程中从用户行为中学习，以复制人类与生俱来的将形状描绘适应上下文特定局部信息的能力，是图像理解研究的一个重要领域。当前的图像分割方法通常采用特定设计，要么依赖通用图像特征，要么依赖特定先验知识，这通常会阻碍它们在不同上下文环境中的应用。本文提出了一种基于强化学习的通用分割框架。它展示了如何就地吸收用户特定行为以实现有效的模型适应和学习。它包含一个两层强化学习算法，该算法根据用户交互过程中积累的经验构建模型。由于该算法在用户进行手动分割时“全面地”学习，因此训练过程无需额外步骤，随着经验的积累，该方法能够自适应并提高其准确性。在体内磁共振（MR）数据上对该方法进行的详细验证表明，该技术在显著降低所需用户交互水平的同时，保持整体分割准确性方面具有实际价值。
