A Low-Resource Training Strategy for Cell Segmentation using Patch-Based Attention U-Net
Keywords: Low-resource learning, Cell segmentation, Microscopy imaging, Custom loss
TL;DR: A low-resource and fine-tuning-free alternative to state-of-the-art cell segmentation models, capable of handling diverse ROI morphologies.
Abstract: Segmentation is an essential tool for cell biologists and involves isolating cells or cellular features from microscopy images. An automated segmentation pipeline with high precision and accuracy can significantly reduce manual labor and subjectivity. Frequently, researchers would seek for a validated model available online and fine-tune it to meet their segmentation requirements. However, the established fine-tuning approach may involve online training or computationally intensive offline training. To address this, we propose an offline training pipeline requiring only tens of samples that are morphologically distinct from pre-training data. Specifically, we employed a patch-based attention U-Net trained with a threshold-based custom loss function. Finally, we evaluated this workflow along with two other state-of-the-art models, Stardist and Cellpose, on three different tasks. Our method improves image segmentation performance by 32.60\% and 35.62\% over Stardist and Cellpose, respectively, using the same amount of training samples.
Submission Number: 16
Loading