LIVECell Dataset for Cell Segmentation
- LIVECell is a publicly available dataset featuring annotated live-cell microscopy images with precise segmentation masks.
- The dataset includes 3,180 phase-contrast images from eight human cell lines captured under diverse conditions and tiled into 256×256 patches.
- It serves as a benchmark for validating deep learning segmentation models, demonstrating robust performance in cross-modality generalization.
The LIVECell dataset is a publicly available collection of annotated live-cell microscopy images designed to support high-precision segmentation and quantitative analysis of human cell cultures. Sourced from Sartorius, LIVECell is widely used as an external testbed for deep learning-based segmentation approaches, particularly for validating the generalization of models trained on alternate modalities such as bright-field microscopy. In contemporary research, LIVECell's comprehensive annotation and challenging imaging characteristics have positioned it as a standard for benchmarking instance and semantic segmentation algorithms in biomedical image analysis (Das et al., 17 Aug 2025).
1. Dataset Composition and Structure
The LIVECell dataset comprises 3 180 phase-contrast microscopy images, each of 2 048 × 2 048 pixels in 8-bit PNG format. The images cover eight human cell lines: A172, SKOV3, SkBr3, MCF7, BV2, BT474, Huh7, and SHSY5Y. For cross-condition diversity, these lines are captured under 22 distinct well/time-point settings, denoted by file name encodings such as “SKOV3_Phase_G4_1_01d04h00m_3.png.” All images used in the referenced external validation were exclusively phase-contrast; no bright-field variants from LIVECell were included. Out of 3 188 originally downloaded frames, eight corrupted files were excluded, yielding a pure test set with no subsequent train-validation split (Das et al., 17 Aug 2025).
For computational efficiency and full-resolution inference, each 2 048 × 2 048 image is tiled into non-overlapping 256 × 256 patches. Predictions from these patches are reassembled to reconstruct the output mask at native resolution.
2. Annotation Schema and Metadata
LIVECell provides high-fidelity instance segmentation masks accompanying every frame, encoded as 8-bit PNG format ground-truths. In the cited study, these were mapped to binary semantic masks, with pixel values of 1 representing cell foreground and 0 denoting background. The segmentation was strictly cell-versus-background, irrespective of instance label. File-level metadata includes details of cell line, imaging modality, plate ID, well coordinate, and timepoint, all embedded in the filename. The external evaluation did not utilize instance identifiers or fields beyond this metadata; assessment centered on mask-based pixel-level and object-level measures (Das et al., 17 Aug 2025).
3. Dataset Statistics and Image Characteristics
Although the average per-frame cell count is not specified in the referenced work, the LIVECell repository reports a nominal range of 10–40 cells for phase-contrast images. The dataset exhibits well-known imaging challenges such as uneven illumination profiles, low interfacial contrast at cell boundaries, moderate background debris, and intermittent motion blur, which reflect typical conditions for live-cell microscopy. The study does not report quantitative metrics for SNR or CNR for LIVECell. Augmentation protocols—such as blur, noise, elastic distortion, random crops, flips, and brightness/contrast jitter—were applied only to the proprietary bright-field training data; LIVECell test images were subjected solely to min-max normalization post-tiling, without further augmentation (Das et al., 17 Aug 2025).
4. Licensing, Distribution, and Access
LIVECell is publicly accessible for academic research purposes under Sartorius’ stipulated non-commercial terms. Users may register for direct download at https://sartorius-research.github.io/LIVECell/. The original images, masks, and associated metadata are distributed in standardized formats to facilitate reproducibility and third-party evaluation. In the cited work, the authors make their complementary bright-field dataset and inference code available on reasonable request, further supporting external validation workflows (Das et al., 17 Aug 2025).
5. Evaluation Protocols and Performance Metrics
External validation on LIVECell uses classical segmentation metrics at both pixel and region overlap levels. These include pixel-wise accuracy:
Precision (positive predictive value), recall (sensitivity), F₁-score, intersection-over-union (IoU/Jaccard), Dice coefficient, as well as SSIM and Hausdorff distance (the latter two not detailed in the summary). Segmentation masks are thresholded to binary form for metric calculations against the ground-truth annotations. The evaluation uses no test-time augmentation beyond normalization and patch assembly (Das et al., 17 Aug 2025).
The following table summarizes key reported metrics for the cited model evaluated on LIVECell (N = 3 180):
| Metric | Value (mean ± std) | Additional Notes |
|---|---|---|
| Accuracy | 0.93 ± 0.06 | Pixel-wise |
| Precision | 0.84 ± 0.11 | TP/(TP+FP) |
| Recall | 0.96 ± 0.04 | TP/(TP+FN) |
| F₁-Score | 0.89 ± 0.07 | 2·(Prec·Rec)/(Prec+Rec) |
| IoU | 0.81 ± 0.10 | TP/(TP+FP+FN) |
| SSIM | 0.99 ± 0.01 | Structural similarity index |
| Hausdorff | 59.2 px ± 36.9 px |
Statistical procedures included a one-sample t-test (t = 112.90, p < 10⁻²⁸⁰) against a baseline F₁ = 0.75, with Cohen’s d of 2.0, demonstrating >99.999% confidence that the model's F₁-score markedly exceeds this threshold (Das et al., 17 Aug 2025).
6. Applications and Impact
LIVECell has become the benchmark for evaluating cross-modality generalization of segmentation pipelines, especially for models trained on bright-field data but required to perform on phase-contrast imagery. In the referenced study, a CNN-based model trained primarily with bright-field images, and with limited (<10%) phase-contrast exposure, achieved robust performance on LIVECell, with 0.89 ± 0.07 F₁-score across eight cell lines and 22 condition strata. This underscores LIVECell’s value for quantifying a model’s robustness, domain adaptation capability, and readiness for deployment in diverse laboratory environments. The absence of test-time augmentation, combined with strong statistical outcomes, attests to the representativeness and rigor of LIVECell as an external validation standard (Das et al., 17 Aug 2025).