Biological processes can be observed both in space and over time using imaging. Visual assessment becomes limiting as datasets grow, and complexity of data as well as subtleness
of processes makes it difficult to draw confident conclusions without automated and quantitative measurement strategies. Traditionally, digital image processing has relied on engineering mathematical models of e.g. the size and shape of cell nuclei, surrounding cytoplasm and fluorescent signal distributions, to extract measurements and apply classification strategies. These methods are powerful, but they are also limited by how well we manage to find a good set of feature descriptors for what we observe. In the past ten years, learning-based approaches relying on deep convolutional neural networks (DCNNs) have gained enormous popularity in all fields of image- based science. The methods have great potential, but they also require care in their usage, where again, the traditional image processing methods play crucial role. We develop and apply combinations of traditional and learning-based methods in a range of areas of systems biology: We apply DCNNs in understanding of cell and bacterial dynamics, drug screening in model organisms, and in decoding and exploration of spatially resolved gene expression in tissue and show results on segmentation of time-lapse imaging of unstained cells, unsupervised feature extraction, and learning of localized gene expression patterns.