Whole-cell segmentation of tissue images with human-level performance using large-scale data annotation and deep learning
read more
Citations
Signatures of plasticity, metastasis, and immunosuppression in an atlas of human small cell lung cancer
Spatial components of molecular tissue biology
Multiplexed imaging mass cytometry of the chemokine milieus in melanoma characterizes features of the response to immunotherapy
Image-based cell phenotyping with deep learning.
Strategies for Accurate Cell Type Identification in CODEX Multiplexed Imaging Data.
References
Deep Residual Learning for Image Recognition
U-Net: Convolutional Networks for Biomedical Image Segmentation
NIH Image to ImageJ: 25 years of image analysis
Fiji: an open-source platform for biological-image analysis
Scikit-learn: Machine Learning in Python
Related Papers (5)
Frequently Asked Questions (13)
Q2. What is the final instance segmentation mask for each nucleus and cell in the image?
The centroid and boundary predictions are used as inputs to a watershed algorithm58 to create the final instance segmentation mask for each nucleus and each cell in the image (Methods).
Q3. How did the authors integrate the morphological features that the authors extracted?
To integrate the morphological features that the authors extracted, the authors performed k-means clustering on the morphology profiles that the authors collected for every cell (Methods).
Q4. Why was annotation time linked to model performance?
Because annotators only needed to correct the mistakes made by the model, not annotate every cell in each image, annotationtime was linked to model performance.
Q5. What is the threshold for the intersection over union of the cell masks?
The authors use the intersection over union (IOU) of the cell masks as the criterion for assessing whether two cells match, with thresholds of 0.4 and 0.1 for the cost matrix and IOU overlaps, respectively.
Q6. What is the main reason for the inaccuracies in the segmentation of cells?
Inaccuracies in segmentation can lead to substantial bias in the identification and enumeration of the cells present in an image.
Q7. How did the authors simulate low signal-to-noise ratio and high background staining?
To simulate low signal-to-noise ratio and high background staining, the authors added uniform random noise of increasing magnitude to each pixel.
Q8. How did the authors compute the localization of a panel of markers with known profiles?
For each channel, the authors selected fields of view in which the marker showed clear expression, and computed the localization within each cell, after removing the bottom 20% lowly expressing cells within each marker.
Q9. Why did Cellpose fail to identify a large fraction of the cells in the image?
In line with its lower recall score (Figure 2c), Cellpose failed to identify a large fraction of the cells in the image (Figure 2f), likely due to the relative scarcity of tissue images in the data used to train Cellpose.
Q10. How did the authors combine the dataset used for model training?
To construct the dataset used for model training, individual .npz files containing annotated images from a single experiment were combined.
Q11. How did the authors find that the crowd annotation of dense images was better?
To support accurate crowd annotation of dense images, the authors found that supplying smaller image crops led to significantly better crowd performance (data not shown).
Q12. Why is the border of each cell in an image to be manually demarcated?
This limitation is largely due to the linear, time-intensive approach used to construct them, which requires the border of every cell in an image to be manually demarcated.
Q13. What are some recent tools that have attempted to overcome this barrier?
Several recent tools have sought to overcome this barrier with a variety of software-engineering approaches, including browser-based software (ImJoy41), Google Colab (ZeroCostDL4Mic42), a centralized web portal (NucleAIzer29, Cellpose28, DeepCell39), and ImageJ plugins (StarDist60, DeepCell39).