YeastSpotter: accurate and parameter-free web segmentation for microscopy images of yeast cells.

Our underlying segmentation method is based upon transferring publicly available convolutional neural networks from the 2018 Kaggle Data Science Bowl competition. In this competition, contestants trained models to segment images of mostly human nuclei, using image set BBBC038v1 from the Broad Bioimage Benchmark Collection ( Ljosa et al. , 2013 ) as training data. Despite not being trained on yeast cells, we found that these models transferred well without fine-tuning. We used a pre-trained mask-RCNN model ( He et al. , 2017 ) by the third-place winner, the Deep Retina team, which we chose due to its simplicity and easily extensible code. To make this model more accessible to the community, we implemented YeastSpotter as a web application to run images through this model. To use YeastSpotter, the user simply uploads their image, which redirects them to a page that tracks the progress of their request and produces segmentation results once ready. A preview image on the result page shows the outlines of the segmentation overlaid on the original input. The user can then download the segmentation, which is stored as an integer-signed tiff file (pixels with a value of 0 correspond to the background, while pixels belonging to each unique cell are each assigned a different integer value). On the website, we provide instructions for loading these fines into ImageJ and scripts to read them in Python, Matlab and R. YeastSpotter is intended for low-throughput use and only accepts a single image per request. For batch segmentation, we also provide user-friendly Python code ( www.github.com/alexxijielu/yeast_segmentation ).

Supplementary Material btz402_Supplementary_Data Click here for additional data file.