Background Root segmentation is a fundamental yet challenging task in image-based plant phenotyping. Accurate segmentation is a prerequisite for extracting root traits relevant to plant physiology, breeding, and agronomy. While U-Net and other convolutional neural network (ConvNet) architectures have been applied to root segmentation, no systematic comparison of multiple Transformer and ConvNet architectures has been conducted across diverse root imaging conditions. Results We evaluated 21 segmentation architectures across nine diverse root image datasets, training 1511 models to assess all combinations of architecture, dataset, pre-training strategy, and learning rate, producing over 3 million segmentations for evaluation. Transformer-based models significantly outperformed ConvNets for Dice (mean Dice 0.679 vs 0.659; <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(p = 3.0 \times 10^{-3}\)</EquationSource> </InlineEquation>). Root-diameter and root-length correlation were also higher for Transformers, but the differences were not statistically significant (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(p = 0.054\)</EquationSource> </InlineEquation> and <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(p = 0.198\)</EquationSource> </InlineEquation> respectively). Pre-training significantly improved mean Dice from 0.623 to 0.666 (<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(p = 6.6 \times 10^{-10}\)</EquationSource> </InlineEquation>), with Transformers benefiting more from pre-training than ConvNets (Dice improvement + 0.072 vs + 0.021; <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(p = 3.7 \times 10^{-4}\)</EquationSource> </InlineEquation>), supporting the hypothesis that fine-tuned Transformers transfer more effectively across large domain gaps. MobileSAM achieved the highest Dice score (0.693) while maintaining computational efficiency. Both architecture families underestimated thin root length compared to manual annotations. Dataset choice explained 70.9% of performance variance, far exceeding model architecture (6.7%). Purpose Transformer architectures significantly outperform ConvNets for root segmentation accuracy, and pre-training significantly improves performance, particularly for Transformers. Pre-trained MobileSAM offers the best accuracy at competitive computational cost. Dataset choice dominates performance variance, suggesting practitioners should prioritize data curation over architecture selection.

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

期刊论文
出版模式：

开放获取
发表日期：

2026年4月20日

查看全文

Abraham George Smith,
Sotiris Lamprinidis,
Anand Seethepalli,
Larry M. York,
Eusun Han,
Patrick Möhl,
Kyriaki Boulata,
Kristian Thorup-Kristensen,
Jens Petersen

摘要

Background

Root segmentation is a fundamental yet challenging task in image-based plant phenotyping. Accurate segmentation is a prerequisite for extracting root traits relevant to plant physiology, breeding, and agronomy. While U-Net and other convolutional neural network (ConvNet) architectures have been applied to root segmentation, no systematic comparison of multiple Transformer and ConvNet architectures has been conducted across diverse root imaging conditions.

Results

We evaluated 21 segmentation architectures across nine diverse root image datasets, training 1511 models to assess all combinations of architecture, dataset, pre-training strategy, and learning rate, producing over 3 million segmentations for evaluation. Transformer-based models significantly outperformed ConvNets for Dice (mean Dice 0.679 vs 0.659; \(p = 3.0 \times 10^{-3}\) ). Root-diameter and root-length correlation were also higher for Transformers, but the differences were not statistically significant ( \(p = 0.054\) and \(p = 0.198\) respectively). Pre-training significantly improved mean Dice from 0.623 to 0.666 ( \(p = 6.6 \times 10^{-10}\) ), with Transformers benefiting more from pre-training than ConvNets (Dice improvement + 0.072 vs + 0.021; \(p = 3.7 \times 10^{-4}\) ), supporting the hypothesis that fine-tuned Transformers transfer more effectively across large domain gaps. MobileSAM achieved the highest Dice score (0.693) while maintaining computational efficiency. Both architecture families underestimated thin root length compared to manual annotations. Dataset choice explained 70.9% of performance variance, far exceeding model architecture (6.7%).

Purpose

Transformer architectures significantly outperform ConvNets for root segmentation accuracy, and pre-training significantly improves performance, particularly for Transformers. Pre-trained MobileSAM offers the best accuracy at competitive computational cost. Dataset choice dominates performance variance, suggesting practitioners should prioritize data curation over architecture selection.