Evaluation of Deep Learning Strategies for Nucleus Segmentation in Fluorescence Images

Abstract

Identifying nuclei is often a critical first step in analyzing microscopy images of cells, and classical image processing algorithms are most commonly used for this task. Recent developments in deep learning can yield superior accuracy, but typical evaluation metrics for nucleus segmentation do not satisfactorily capture error modes that are relevant in cellular images. Besides, large image data sets with ground truth for evaluation have been limiting. We present an evaluation framework to measure accuracy, types of errors, and computational efficiency; and use it to compare two deep learning strategies (U-Net and DeepCell) alongside a classical approach implemented in CellProfiler. We publicly release a set of 23,165 manually annotated nuclei and source code to reproduce experiments. Our results show that U-Net outperforms both pixel-wise classification networks and classical algorithms. Also, our evaluation framework shows that deep learning improves accuracy and reduces the number of biologically relevant errors by half.

Test-time augmentation for deep learning-based cell segmentation on microscopy images

“…Most of the images have come from the stage 1 train/test data of Data Science Bowl 2018. We also used additional sources [14][15][16][17][18][19][20] and other data published in the discussion thread 'Official External Data Thread' (https://www.kaggle.com/c/ data-science-bowl-2018/discussion/47572) related to DSB 2018. The images were labelled by experts using the annotation plugins of ImageJ/Fiji and Gimp.…”

A deep learning framework for nucleus segmentation using image style transfer

“…DSB test1 and DSB test2 are heterogeneous test sets from the Kaggle challenge (stage 1 and stage 2). The fluo dataset is fluorescence images of U2OS cells in a chemical screen taken from the Broad Bioimage Benchmark Collection (BBBC039) 9 . The hist dataset is a mixture of histology images collected from the internet and prostate H&E stained slides collected in-house.…”

“…It is also responsible for breakthroughs in diagnosing retinal images 6 , classifying skin lesions with superhuman performance 7 , as well as incredible advances in 3D fluorescence image analysis 8 . However, aside from initial works from Caicedo et al 9 and Van Valen et al 10 , deep learning has yet to significantly advance nucleus segmentation performance.…”

NuSeT: A Deep Learning Tool for Reliably Separating and Analyzing Crowded Cells

“…10 The final foreground prediction is then computed from the maximum class score of each pixel. Although U-Net alone performs well on some microscopy datasets 22,26 , we incorporated RPN since it was originally designed to detect objects in images with high information content. 17 We reasoned that the accurate performance of RPN in detecting objects can be leveraged to improve nuclear segmentation performance.…”

“…17,18 Additionally, at the pixel level, the segmentation task of Mask-R-CNN is performed by FCN, which is less accurate with small training datasets compared with U-Net. 15,22 To address these issues, we have developed a Nuclei Segmentation Toolset (NuSeT), which integrates U-Net 15 and a modified RPN (based on the implementation of previous works 23,24 ) to accurately segment fluorescently labeled nuclei. In this integrated model, U-Net performs pixelsegmentation, while the modified RPN predicts unique bounding boxes for each image based on U-Net segmentations.…”