← Papers

Unverified paper record

Transformers Outperform ConvNets for Root Segmentation: A Systematic Comparison Across Nine Datasets

bioRxiv (Cold Spring Harbor Laboratory) · 19 Feb 2026 · 10.64898/2026.02.18.706562

Abstract

Abstract Root segmentation is a fundamental yet challenging task in image-based plant phenotyping. We present the first systematic comparison of Transformer and Convolutional Neural Network (ConvNet) architectures for root segmentation, evaluating 21 architectures across nine diverse datasets and comparing pre-trained models to training from scratch. Transformer-based models significantly outperform ConvNets for segmentation accuracy and root-diameter agreement. Pre-training significantly improves mean Dice from 0.623 to 0.666 ( p = 3.3 × 10 −10 ). We also find that Transformers benefit more from pre-training than ConvNets, with Dice improvements of +0.072 versus +0.022 ( p = 3.7 × 10 −4 ), supporting the hypothesis that fine-tuned Transformers transfer more effectively across large domain gaps. Among evaluated models, MobileSAM achieved the highest Dice score while maintaining computational efficiency. Dataset choice explained far more performance variance (70.9%) than model architecture (6.7%), suggesting that data curation matters more than model selection.

Plant phenotyping relevance

根の画像セグメンテーション手法を21種類・9データセットで体系比較し、植物フェノタイピングにおける精度と根径推定を検証しているため、方法評価が中心である。

abstractRoot segmentation is a fundamental yet challenging task in image-based plant phenotyping.
abstractWe present the first systematic comparison of Transformer and Convolutional Neural Network (ConvNet) architectures for root segmentation, evaluating 21 architectures across nine diverse datasets
abstractcomparing pre-trained models to training from scratch

Code and data availability

The paper's authors explicitly state that their training/segmentation analysis code is publicly available on GitHub. The nine root image datasets evaluated are cited prior public datasets (DeepRootLab, Grassland, Chicory, PRMI), not paper-specific assets of this study, so the authors' own code repository is the only in

Codepublic

r of parameters, as these affect hardware requirements, running costs, and environmental impact. To jointly compare efficiency and accuracy, we ranked models by the mean of their Dice, parameter count, and FLOPs ranks, providing a simple combined metric for practitioners balancing these trade-offs. Training code is available at https://github.com/sotlampr/seg.Configuration selection To prevent overfitting to the test set, model selection used a two-stage procedure based on validation performance: Replicate selection: For each combination of model, dataset, learning rate, and pre-training, the replicate with the highest validation Dice was retained, along with its paired test result. Hyperpar

Open resource ↗sotlampr/seg · pdf-raw-page:4 lines:1-95

This is an automatically classified, unverified record. Curator approval is required before any resource enters the Catalog.