The story behind the root image library
Graphical abstract of our paper
Some days ago, our manuscript, Using a Structural Root System Model to Evaluate and Improve the Accuracy of Root Image Analysis Pipelines was published online in its final version. The road has been long to reach that point, but it was a very fruitful one. And it was, in my opinion, a perfect example of the utility of both preprints and peer-review.
Let me explain.
At the beginning, the paper was built around one simple idea: we can use structural plant models (here, ArchiSimple) to generate large libraries of root system images, together with their corresponding ground-truth (e.g., for roots, the true number of roots in the system). Such libraries can then be used to validate and calibrate image analysis tools.
Preprint accelerated dissemination and triggered early comments
At that time, in Septembre 2016, we submitted the paper to the Frontiers in Plant Science special issue on Plant morphological modeling](http://journal.frontiersin.org/researchtopic/4300). At the same time, we deposited the manuscript on bioRxiv. It was therefore directly available for the research community to read and comment. It was quickly viewed more than 1000 times and the feedback we received was important, including a full twitter review by Larry York and Ruben Rellan Alvarez.
Among other things, this informal review allowed us to already identify where the manuscript was confusing and the terms ambiguous. In particular, we introduced the concept of synthetic genotype, or synthetype that was not clear enough. Ultimately, we even removed it from te paper...
Peer-review greatly improved the quality of our paper
A few weeks later, the reviewers report came back from Frontiers. The reviewers (Tony Pridmore and Chris Topp) made a lot of very valid points and criticisms. They pointed some important limitations (and mistakes) in our initial manuscript. As a result, we ended up making a lot of modification the manuscript, (v1 here, v2 here).
Here is a couple of examples of things we changed. There was a bug in the image analysis pipeline we used that resulted in strange results that we did not see in the first place. The reviewers also pointed out that we used only perfect images, which is somehow unrealistic. So we generated three parallel datasets, with different level of noise. The most important change was the addition of a whole new section. The reviewers rightfully pointed out that we could use the synthetic image library to calibrate root image analysis tools. We took it a step further and showed that we could use our synthetic library to train machine learning algorithms (in this case Random Forest). The machine learning approach allowed us to obtain much better estimation of root traits (see figure below), even for very large root systems. Our manuscript was much better in its second and third versions.
Example of results obtained with (right) and without (left) the machine learning approach. With the machine learning approach, the estimation of the total number of roots was much better, in particular for large root systems
Preprint and peer-review work great together
To conclude, for me, this manuscript will then stay as a perfect example of the use of both preprint and peer-review. The preprint shortened the dissemination of our work by seven months (Sept 2016 vs April 2017), while the peer-review process greatly improved it. A win-win for us.
- Preprint: https://doi.org/10.3389/fpls.2017.00447
- Paper: https://doi.org/10.1101/074922
- Root image library: http://doi.org/10.5281/zenodo.208214
- Code used in the paper: http://doi.org/10.5281/zenodo.208499
- Machine learning code: https://github.com/FaustFrankenstein/RandomForestFramework/releases/tag/v1.0
blog comments powered by Disqus