A systematic comparison of transformers and ConvNets for root segmentation across nine datasets

Area of Science:

• Plant Science
• Computer Vision
• Bioinformatics

Background:

• Root segmentation is crucial for plant phenotyping, impacting plant physiology, breeding, and agronomy.
• Existing studies often use Convolutional Neural Networks (ConvNets) like U-Net, but lack systematic comparisons with Transformer architectures.
• Diverse root imaging conditions present challenges for accurate segmentation.

Purpose of the Study:

• To systematically compare Transformer and ConvNet architectures for root segmentation accuracy across diverse datasets.
• To evaluate the impact of pre-training strategies on segmentation performance.
• To identify the best-performing model and assess the influence of dataset choice versus model architecture on performance.

Main Methods:

• Evaluated 21 segmentation architectures (including Transformers and ConvNets) across nine diverse root image datasets.
• Trained 1511 models, exploring combinations of architecture, dataset, pre-training, and learning rate.
• Generated over 3 million segmentations for comprehensive evaluation using metrics like Dice score.

Main Results:

• Transformer models significantly outperformed ConvNets in Dice scores (0.679 vs 0.659).
• Pre-training boosted mean Dice scores significantly (0.623 to 0.666), with Transformers benefiting more.
• Pre-trained MobileSAM achieved the highest Dice score (0.693) with computational efficiency; dataset choice (70.9%) had a greater impact than architecture (6.7%).

Conclusions:

• Transformer architectures offer superior accuracy for root segmentation compared to ConvNets.
• Pre-training is a vital strategy, particularly for enhancing Transformer performance across domain gaps.
• Prioritizing data curation over architecture selection is recommended for optimizing root segmentation in plant phenotyping.