Labkit: Labeling and Segmentation Toolkit for Big Image Data
read more
Citations
Active mesh and neural network pipeline for cell aggregate segmentation
References
Random Forests
NIH Image to ImageJ: 25 years of image analysis
Fiji: an open-source platform for biological-image analysis
The WEKA data mining software: an update
Ror2 signaling regulates Golgi structure and transport through IFT20 for tumor invasiveness
Related Papers (5)
DeepMIB: User-friendly and open-source software for training of deep learning network for biological image segmentation
Frequently Asked Questions (18)
Q2. What have the authors stated for future works in "Labkit: labeling and segmentation toolkit for big image data" ?
In the future, the authors intend to extend LABKIT ’ s functionalities to improve manual and automated segmentation. For instance, the authors will add a magic wand tool to select, fill, fuse or delete labels based on the pixel classification. LABKIT source code is open source and can be found online ( 49 ), together with its command-line interface ( 36 ) and tutorials ( 50 ).
Q3. What can be done to create a segmentation mask?
Once the image is fully segmented, the generated segmentation masks can be transferred to label layers and the drawing tools can now be used to curate them.
Q4. What can be used to curate segmentation results obtained by other methods?
LABKIT can also be used to curate segmentation results obtained by other methods that are not available within LABKIT, including deep learning based methods (34).
Q5. What is the use of sparse labeling in LABKIT?
sparse labeling combined with random forest pixel classification in LABKIT was used to segment mice epidermal cells (32), as well as mRNA foci in neurons (33).
Q6. What is the way to generate the ground-truth data?
In cases where higher segmentation quality is truly necessary, curated results from shallow learning can be used to generate the massive amount of ground-truth required to train a deep learning algorithm.
Q7. How many CPU nodes can run LABKIT on a single laptop?
While running computation on an HPC cluster comes with overhead, increasing the number of CPU nodes shortens the computation dramatically, reaching a 40- fold improvement from 1 CPU node to 50.
Q8. What are the latest techniques for analyzing biological tissue?
In recent years, new and powerful microscopy and sample preparation techniques have emerged, such as lightsheet (1), super-resolution microscopy (2–6), modern tissue clearing (7, 8), or serial section scanning electron microscopy (9, 10) enabling researchers to observe biological tissues and their underlying cellular and molecular composition and dynamics in unprecedented details.
Q9. How many GPU nodes can process the image in a minute?
GPU nodes on an HPC allow for more parallelization of the computation and therefore even higher computational speed-up on the segmentation task, with 10 GPU nodes processing the data in slightly over a minute.
Q10. What is the main purpose of LABKIT?
LABKIT features a user-friendly interface allowing for rapid scribble labeling, training, and interactive curation of the segmented image.
Q11. What are the foundations for a software tool?
The required foundations for such a software tool have in recent years been built by the vibrant research software engineering community around Fiji and ImageJ2.
Q12. What is the main reason why the Fijian community is not capable of processing large datasets?
It is, regrettably, not capable of processing very large datasets due to its excessive demand for CPU memory, leaving the sizable Fiji community with a lack of user-friendly pixel classification or segmentation tools that can operate on large multi-dimensional data.
Q13. What are examples of techniques used to perform deep learning?
Examples for such methods range from intensity thresholding and seeded watershed, to shallow machine learning approaches on manually chosen or designed features.
Q14. What can be used for more complex segmentation tasks?
For more complex segmentation tasks that need to discriminate various visible structures (e.g. nucleus vs. cytoplasm vs. background) or cell types (as in Fig. 2C), two or more foreground classes can be used (Fig. 2D).
Q15. How can LABKIT be used to simplify the integration into existing workflows?
to simplify the integration into existing workflows in Fiji, LABKIT can be easily called from the ImageJ macro language.
Q16. What is the classification method for large images?
LABKIT pixel classification ranked as the highest performing segmentation method on this dataset for all three evaluation metrics (OPCSB , SEG and DET ) (40).
Q17. How long did it take to segment all images?
In contrast, manually segmenting all images required more than an hour (Fig. 4D), which is four times longer than scribble-based pixel classification with LABKIT, followed by full curation of the results to obtain images of comparable quality.
Q18. What is the way to generate ground-truth data for a deep learning method?
Generating ground-truth data for a deep learning method is a tedious endeavour without the insurance of a perfect segmentation result.