Combining semi-automated image analysis techniques with machine learning algorithms to accelerate large-scale genetic studies

Abstract

Genetic analyses of plant root systems require large datasets of extracted architectural traits. To quantify such traits from images of root systems, researchers often have to choose between automated tools (that are prone to error and extract only a limited number of architectural traits) or semi-automated ones (that are highly time consuming). We trained a Random Forest algorithm to infer architectural traits from automatically extracted image descriptors. The training was performed on a subset of the dataset, then applied to its entirety. This strategy allowed us to (i) decrease the image analysis time by 73% and (ii) extract meaningful architectural traits based on image descriptors. We also show that these traits are sufficient to identify the quantitative trait loci that had previously been discovered using a semi-automated method. We have shown that combining semi-automated image analysis with machine learning algorithms has the power to increase the throughput of large-scale root studies. We expect that such an approach will enable the quantification of more complex root systems for genetic studies. We also believe that our approach could be extended to other areas of plant phenotyping.

Keywords: QTL analysis; machine learning; plant phenotyping; root.

Figure 1:

Overview of the analysis pipeline used in this study. (A) We divided the full dataset (2614 images) into two: a training set (100 to 900 images) and a test set (1645 images). (B) For each dataset, all the images were analysed using a semi-automated root image analysis tool (RootNav) to extract the ground-truth, as well as with a fully automated root image analysis tool (RIA-J) to extract image descriptors (see the text for details). (C) We trained a Random Forest model on the image descriptors and the ground-truth from the training dataset. (D) We applied the Random Forest model on the image descriptors from the test dataset. (E) We compared the image descriptors and the Random Forest estimators from the test dataset with their corresponding ground-truth. (F) Comparison of biologically relevant metrics extracted with the automated analysis and the Random Forest analysis. (G) QTL were identified and compared using both Random Forest estimators and the ground-truth data.

Figure 2:

Accuracy of the Random Forest estimators. The R² values of the linear regression between the Random Forest estimators and the ground-truths were computed for each size and repetition of training datasets. The dotted line represents the R² value between the most closely related image descriptors and the ground-truth.

Figure 3:

Screenshot of PRIMAL. (A) Variable to evaluate with the Random Forest algorithm. (B) Random Forest algorithm parameters. (C) Visualization of the accuracy of the Random Forest estimators. (D) Accuracy metrics for the different descriptors.