Learning accurate plant cell segmentation and curation, with minimal human effort

快速可靠的植物细胞分割是FOR 2581合作者进行大量定量工作的先决条件。在上一个资助期内，我们成功开发了一种结合“深度学习”和“结构化预测”的方法，可以在体积光学显微镜数据中以非常高的精度分割植物细胞。不幸的是，这些方法需要大量的训练数据才能可靠地工作。虽然人类擅长分析自然的2D或2.5D图像，但他们没有天生的能力来有效地分析或处理真正的3D图像数据。因此，考虑到分割任务本身的最新进展，体积图像分析的主要瓶颈已经转移到创建用于训练的3D地面实况，以及由人类专家管理算法结果。作为回应，该项目有两个主要目标：首先，进一步开发我们的分割工具，使它们可以仅使用最少的注释进行训练;其次，开发机器学习方法，可以发现并在可能的情况下自动纠正可能的分割错误，也是用最少的人工注释进行训练。我们将联合收割机的开发整合到一个通用管道中，该管道将从多个薄弱、易于提供的监督来源提供高度准确的分割。正如在第一个资助期，新的方法和想法将不断面对来自研究单位的真实的数据，以建立优点和局限性。表现最好的方法将被传递到中心项目Z 02，该项目将把研究代码转化为RU合作伙伴和整个生物成像社区的开源最终用户工具。

A fast and reliable segmentation of plant cells is a prerequisite for much of the quantitative work pursued by the collaborators of FOR2581. In the last funding period, we have successfully developed methods – a combination of “deep learning” and “structured prediction” – that allow segmenting plant cells in volumetric light microscopy data with very high accuracy. Unfortunately, these methods require a lot of training data to work reliably. While humans excel in the analysis of natural 2D or 2.5D images, they have no innate ability to efficiently analyze or process truly 3D image data. As a consequence and given the latest advances on the segmentation task itself, the main bottleneck of volumetric image analysis has shifted to the creation of 3D ground truth for training, and to curation of the algorithm results by a human expert.In response, this project has two principal aims: first, to further develop our segmentation tools such that they can be trained using minimal annotations only; and secondly, to develop machine learning methods that can spot and, if possible, automatically correct likely segmentation errors, also trained with minimal human annotation. We will combine our developments in a single generic pipeline that will deliver highly accurate segmentations from multiple sources of weak, easily provided supervision. Just as in the first funding period, the new methods and ideas will be constantly confronted with real data from the research unit to establish virtues and limitations. The best-performing methods will be passed to central project Z02 which will turn research code into open source end-user tools for RU collaboration partners and the bioimaging community at large.