Figure 2
From: Generalising from conventional pipelines using deep learning in high-throughput screening workflows
Qualitative analysis of the generalisation capability of CDL vs CIP. (a) On the left, a sample raw image from the dataset of fluorescent microscopy images (microscopy channels one and two are displayed as red and green channels of the RGB image). On the right, the corresponding (weak) mask is produced with the conventional image processing (CIP) pipeline. The different types of errors are shown in the legend. (b) Examples comparing CIP and CNN based segmentation (CDL) on previously unseen data. Top row: part of original HTS image, middle row: CIP Segmentation, bottom row: CDL segmentation. Errors of the segmentation are highlighted by black arrows. Note the improvements the CNN learned over the CIP which was used to generate training data.
