DeepFLaSH, a deep learning pipeline for segmentation of fluorescent labels in microscopy images

Abstract Here we present and evaluate DeepFLaSH, a unique deep learning pipeline to automatize the segmentation of fluorescent labels in microscopy images. The pipeline allows training and validation of label-specific convolutional neural network (CNN) models that can be uploaded to an open-source CNN-model library. As there is no ground truth for fluorescent signal segmentation tasks, we evaluated the CNN with respect to inter-coding reliability. Similarity analysis showed that CNN-predictions highly correlated with segmentations by human experts. DeepFLaSH also allows adaptation of pretrained, label-specific CNN-models from our CNN-model library to new datasets by means of transfer learning. We show consistent model-performance on datasets from three independent laboratories after transfer learning, thus ensuring its objectivity and reproducibility. DeepFLaSH runs as a guided, hassle-free open-source tool on a cloud-based virtual notebook with free access to high computing power and requires no machine learning expertise.