CellSeg1: Robust Cell Segmentation with One Training Image

TL;DR

CellSeg1 tackles the annotation bottleneck in cell segmentation by harnessing the Segment Anything Model (SAM) with Low-Rank Adaptation (LoRA) to enable robust segmentation from a single, carefully annotated image. The method achieves an average $mAP_{0.5}$ of $0.81$ across 19 diverse datasets and demonstrates strong cross-dataset generalization, notably on TissueNet, while requiring only a few dozen annotations. A key finding is that annotation quality outweighs quantity, with densely packed, multi-scale cells in a single image sufficing for competitive performance. The approach emphasizes practical accessibility, achieving high performance on consumer hardware with a user-friendly GUI and a streamlined training/inference pipeline, offering a pragmatic alternative to universal, data-hungry segmentation models.

Abstract

Recent trends in cell segmentation have shifted towards universal models to handle diverse cell morphologies and imaging modalities. However, for continuously emerging cell types and imaging techniques, these models still require hundreds or thousands of annotated cells for fine-tuning. We introduce CellSeg1, a practical solution for segmenting cells of arbitrary morphology and modality with a few dozen cell annotations in 1 image. By adopting Low-Rank Adaptation of the Segment Anything Model (SAM), we achieve robust cell segmentation. Tested on 19 diverse cell datasets, CellSeg1 trained on 1 image achieved 0.81 average mAP at 0.5 IoU, performing comparably to existing models trained on over 500 images. It also demonstrated superior generalization in cross-dataset tests on TissueNet. We found that high-quality annotation of a few dozen densely packed cells of varied sizes is key to effective segmentation. CellSeg1 provides an efficient solution for cell segmentation with minimal annotation effort.

Figures (11)

• Figure 1: CellSeg1 trained on just a single image outperforms existing methods trained on extensive images. (a) Performance comparison on 5 diverse datasets, where inner radar charts show $mAP_{0.5}$ scores and outer rings show the single training image used for CellSeg1 to achieve the corresponding inner radar chart results. (b) Cross-subset generalization test on 14 TissueNet nuclei subsets, where generalization ability is assessed using leave-one-out cross-validation across subsets.
• Figure 1: Impact of annotation quality and cell count on CellSeg1 performance across diverse datasets. Each row represents a different dataset, with images arranged from left to right by increasing cell count. The top-left of each image shows the $mAP_{0.5}$ achieved on the test set when using that single image for training, while the top-right indicates the number of annotated cells. Yellow outlines in the bottom-right quadrant represent cell annotations.
• Figure 2: The pipeline of CellSeg1 for cell segmentation. The pipeline consists of two main stages: Training (top) and Inference (bottom). During training, the model uses LoRA to fine-tune the SAM on a single point. The inference stage utilizes a grid of points to generate multiple masks, which are then filtered using Non-Maximum Suppression (NMS).
• Figure 2: CellSeg1's performance using single-image training across diverse cell types. (a) High-quality training images, which are the same exemplary images shown in Figure 3g, with their corresponding $mAP_{0.5}$ scores on the test set and cell counts. Yellow outlines indicate ground truth annotations. (b) Test images with predicted cell masks (cyan outlines) and their $AP_{0.5}$ scores. (c,d) Zoomed-in views of regions marked by red and blue boxes in (b), respectively, showing detailed segmentation results.
• Figure 3: Performance analysis of CellSeg1 trained on different single training images. (a-e) Distribution of $mAP_{0.5}$ for each method when trained on different single images from the respective datasets. Boxes centered on medians, whiskers from Q1 - 1.5 IQR to Q3 + 1.5 IQR, constrained to the data range. IQR: interquartile range (Q3-Q1). (f) Scatter plot showing the relationship between the number of cells in each training image and the resulting $mAP_{0.5}$ for CellSeg1. Each point represents a single experiment using one training image. (g) Examples of high-quality training images that led to good performance for CellSeg1. (h) Zoomed-in views of the red boxed areas in (g), highlighting key features of effective training images. (i) Examples of low-quality training images that resulted in poor performance for CellSeg1. (j) Zoomed-in views of the red boxed areas in (i), illustrating characteristics of problematic training images.