RootPainter: deep learning segmentation of biological images with corrective annotation.

Area of Science:

• Plant science
• Computer vision
• Machine learning

Background:

• Convolutional neural networks (CNNs) are valuable for plant image analysis but require machine learning expertise.
• Accessibility of CNNs for researchers without a machine learning background is a significant challenge.

Purpose of the Study:

• To introduce RootPainter, an open-source graphical user interface (GUI) software for rapid deep neural network (DNN) training in biological image analysis.
• To evaluate RootPainter's effectiveness in training models for root length extraction, biopore counting, and root nodule counting.

Main Methods:

• Developed RootPainter, a GUI-based software tool for training DNNs.
• Trained models for specific plant image analysis tasks: root length extraction (chicory), biopore counting, and root nodule counting.
• Compared dense annotations with corrective annotations added during training.

Main Results:

• Models trained with RootPainter using corrective annotations within 2 hours showed strong correlation with manual measurements in 5 out of 6 cases.
• Annotation duration significantly correlated with model accuracy, suggesting further improvements with extended annotation.
• High accuracy was achieved for diverse datasets (varying objects, backgrounds, image quality) with less than 2 hours of annotation time.

Conclusions:

• RootPainter facilitates high-accuracy deep learning model training for plant image analysis within a day.
• The software significantly lowers the barrier to entry for applying machine learning in biological image analysis.
• RootPainter empowers researchers without machine learning backgrounds to perform complex image analysis tasks efficiently.