Digital diatom analysis: investigating advanced deep learning-based approaches for gigapixel-sized virtual slides

硅藻是一类在水生环境中发挥重要生态作用的微藻。对它们的硅酸盐外壳进行光学显微镜研究是最古老的方法之一，但在生态学，古生态学和应用研究中仍然广泛使用，以确定它们的组合成分。这些方法向数字对应物的过渡已经进行了一段时间，在生态/形态学研究中，有大量迹象表明在一致性、透明度、精确度和统计能力方面有所改进。到目前为止，这种数字硅藻分析大多只针对一个或几个类群，或需要大量的人工注释工作的情况下，复杂的社区分类鉴定。然而，试点研究现在已经清楚地表明，深度卷积网络可能很快就能够使数字硅藻分析在“真实的生活”条件下工作，具有丰富的物种群落。我们确定了目前阻碍这一点的三个主要问题领域，涉及：1）尽管背景复杂（由沉积物和其他颗粒引起），但仍对硅藻图像进行分割; 2）面对自然群落中凹秩-丰度关系，有效地增加训练数据集的分类覆盖率; 3）尺寸，轮廓形状和纹理对于算法硅藻识别的相对重要性。我们希望通过一系列深度学习实验来解决这些问题，并为在群落生态学、古生态学和其他类型的硅藻调查中，由深度学习模型支持的数字硅藻分析的常规应用铺平道路。

Diatoms are a speciose group of microalgae playing important ecological roles in a broad range of aquatic habitats. Light microscopic investigation of their silicate shells is one of the oldest, but still widely used approaches to determining their assemblage composition in ecological, paleo-ecological and applied research. A transition of these methods to digital counterparts has been ongoing for some time now, with abundant indications for improvements related to consistency, transparency, precision and statistical power in ecological / morphometric studies. Thus far, such digital diatom analyses have mostly targeted only one or a few taxa, or required substantial manual annotation effort for taxonomic identification in the case of complex communities. Pilot studies have, however, now made it clear that deep convolutional networks will probably soon be able to also enable digital diatom analysis to work in “real life” conditions, with species rich communities. We identified three main problem fields presently hindering this, related to 1) segmentation of diatom images in spite of complex backgrounds (caused by sediment and other particles); 2) efficiently increasing taxonomic coverage of training data sets in the face of concave rank-abundance relationships in natural communities; and 3) the relative importance of size, outline shape and texture for algorithmic diatom identification. We would like to address these problems by a series of deep learning experiments and pave the way for a routine application of digital diatom analysis supported by deep learning models in community ecology, paleoecology and other types of diatom investigations.