A fully automated deep learning pipeline for high-throughput colony segmentation and classification.

Yukiko Shimada,Lea Duempelmann,Marc Bühler,Sarah H. Carl

doi:10.1242/bio.052936

Abstract

ABSTRACTAdenine auxotrophy is a commonly used non-selective genetic marker in yeast research. It allows investigators to easily visualize and quantify various genetic and epigenetic events by simply reading out colony color. However, manual counting of large numbers of colonies is extremely time-consuming, difficult to reproduce and possibly inaccurate. Using cutting-edge neural networks, we have developed a fully automated pipeline for colony segmentation and classification, which speeds up white/red colony quantification 100-fold over manual counting by an experienced researcher. Our approach uses readily available training data and can be smoothly integrated into existing protocols, vastly speeding up screening assays and increasing the statistical power of experiments that employ adenine auxotrophy.

Highlights

Adenine auxotrophy is a commonly used non-selective genetic marker in yeast research
Up to now manual scoring is a common practice in the yeast community
While we were encouraged by our high validation accuracy, we wanted to test the pipeline’s performance against manual counting in a real-world, experimental context

Summary

Background

Auxotrophy is the inability of an organism to synthesize a particular organic compound required for its growth. Two major tasks are required to generate this output: separating the colonies from the plate background and classifying them individually as white or non-white These two tasks could conceivably be completed either in one step, as with a single-shot detector[1] or RetinaNet[2], or in two separate steps, such as with a semantic segmentation, where each pixel is assigned a label such as “foreground” or “background”, followed by classification of cropped images. As insufficient training data is a common problem hampering efforts to apply deep learning in many biological domains[4], we opted to use a pragmatic approach, treating the segmentation and classification steps as separate problems (Fig. 1a) This allowed us to use simpler and, when available, pre-existing annotations for training data: for the segmentation task, we used masks generated previously using the Ilastik image-processing toolkit[5], while for the classification task, we relied on manual labels assigned by experienced biologists to cropped images of single colonies. Segmentation and colony class prediction can be performed separately, allowing for classification of previously-segmented images

Results

Discussion and Conclusions

Methods