Revvity Full Cell Segmentation Dataset

• Revvity-25 is a brightfield whole‐cell instance segmentation dataset featuring high-resolution images with detailed, expert-validated annotations of overlapping cell cytoplasm.
• It includes 110 1080×1080 images with 2937 manually labeled cancer cells, split evenly between training and testing for a compact yet dense benchmark.
• IAUNet, a query-based U-Net architecture, achieves state-of-the-art performance on this dataset by accurately capturing complex cell borders and overlaps.

Searching arXiv for the dataset paper and closely related segmentation benchmarks. The 2025 Revvity Full Cell Segmentation Dataset, also referred to in experiments as Revvity-25, is a brightfield microscopy dataset for cell instance segmentation that was introduced together with IAUNet, a query-based U-Net architecture for biomedical instance segmentation (Prytula et al., 3 Aug 2025). It is defined by detailed annotations of overlapping cell cytoplasm in brightfield images, with a stated emphasis on accurate and detailed annotations for cell borders and overlaps (Prytula et al., 3 Aug 2025). The dataset is positioned as a benchmark for cell instance segmentation, and the paper further states that it “opens new possibilities for testing and benchmarking models for modal and amodal semantic and instance segmentation” (Prytula et al., 3 Aug 2025).

1. Definition and scope

The dataset is presented as a distinct contribution in the IAUNet work: “One of our key contributions in this paper is a novel cell instance segmentation dataset named Revvity-25” (Prytula et al., 3 Aug 2025). Its stated role is to provide detailed, expert-validated instance annotations in brightfield microscopy, specifically for segmentation settings in which overlapping cells and precise cell borders are central difficulties (Prytula et al., 3 Aug 2025).

The task formulation is single-class instance segmentation of whole cells, with the annotations targeting cell cytoplasm rather than nuclei alone. The paper characterizes the target structures as “overlapping cell cytoplasm in brightfield images” and emphasizes “cell borders and overlaps” (Prytula et al., 3 Aug 2025). No separate nucleus labels, multi-class labels, or cell-subtype labels are described. This suggests that the dataset is intended primarily for methods that must distinguish individual cell instances under partial contact or overlap, rather than for multi-structure parsing.

A plausible implication is that Revvity-25 occupies a niche distinct from nucleus-centric fluorescence datasets and from semantic-only cell masks. In the paper’s framing, its importance lies in the combination of brightfield imaging, whole-cell delineation, and high-detail manual contours (Prytula et al., 3 Aug 2025).

2. Image content and dataset composition

The paper gives explicit dataset statistics: “It includes 110 high-resolution 1080 × 1080 brightfield images, each containing, on average, 27 manually labeled and expert-validated cancer cells, totaling 2937 annotated cells” (Prytula et al., 3 Aug 2025). The images therefore constitute a relatively compact but densely annotated benchmark.

The train-test partition is also explicit: “The Revvity-25 dataset is divided equally into train and test sets, each containing 55 images” (Prytula et al., 3 Aug 2025). No separate validation split is described in the paper. The biological material is described only as cancer cells; the paper does not specify the exact cancer cell line or tissue of origin (Prytula et al., 3 Aug 2025).

The dataset’s core reported properties can be summarized as follows:

Property	Reported value
Number of images	110
Image resolution	1080 × 1080
Average cells per image	27
Total annotated cells	2937
Train split	55 images
Test split	55 images
Imaging modality	Brightfield

The paper does not specify objective magnification, numerical aperture, bit depth, file format, or acquisition hardware (Prytula et al., 3 Aug 2025). It also does not describe time series, multiple channels, or live-cell imaging. This suggests a static-image benchmark centered on high-resolution brightfield morphology.

The text further characterizes the morphology as challenging: “the first dataset with accurate and detailed annotations for cell borders and overlaps, with each cell annotated using an average of 60 polygon points, reaching up to 400 points for more complex structures” (Prytula et al., 3 Aug 2025). This indicates substantial contour complexity and implies that at least some objects exhibit irregular boundaries or heavy overlap.

3. Annotation design and label semantics

Revvity-25 is explicitly an instance-level dataset. The paper describes it as containing “hundreds of carefully annotated cell instances in high-resolution brightfield images, each thoroughly hand-labeled and validated” and reports a total of 2937 annotated cells (Prytula et al., 3 Aug 2025). Each annotated object corresponds to a full cell extent, specifically the cytoplasm, rather than only a nucleus.

The annotation granularity is unusually fine for a brightfield benchmark. The paper states that each cell is annotated using an average of 60 polygon points, with up to 400 points for more complex structures (Prytula et al., 3 Aug 2025). These are therefore high-density polygonal contours rather than coarse boxes or low-resolution masks. The stated motivation is accurate representation of cell borders and overlaps.

The annotation process is described as manual and quality-controlled: the cells are “manually labeled and expert-validated” and “thoroughly hand-labeled and validated” (Prytula et al., 3 Aug 2025). No inter-annotator agreement statistic is reported. The body text does not explicitly name the labeling tool, although the references include Label Studio, which strongly suggests that it was used in the annotation workflow (Prytula et al., 3 Aug 2025). Since the paper does not state this directly, that remains an inference.

The representation of overlap is central. The paper emphasizes “precise annotation of cell borders, even in cases of overlapping cells, allowing it to capture complex cell interactions” (Prytula et al., 3 Aug 2025). This suggests that each touching or overlapping cell is delineated as a separate instance, even where image evidence is locally ambiguous. The paper does not explicitly specify whether the release format is COCO polygons, per-instance masks, or another serialization, but because the experiments use Mask R-CNN, Mask2Former, MaskDINO, and IAUNet, a plausible implication is that the annotations are distributed in, or can be converted into, a COCO-like instance format (Prytula et al., 3 Aug 2025).

4. Evaluation protocol and benchmark usage

The dataset is evaluated with COCO-style average precision metrics: AP, AP$_{50}$, AP$_{75}$, AP$_S$, AP$_M$, and AP$_L$ (Prytula et al., 3 Aug 2025). The paper does not restate the formal AP definition, but its evaluation table uses the standard nomenclature and object-size stratification (Prytula et al., 3 Aug 2025).

The experimental preprocessing protocol is given explicitly and applies to Revvity-25 as part of the broader IAUNet evaluation setup. During training, the images undergo longest-side resizing to 512 × 512 pixels, preserving aspect ratio, followed by scale jittering within 0.8 to 1.5, fixed-size cropping to 512 × 512, and random flipping (Prytula et al., 3 Aug 2025). During inference, the paper states that it applies the same resizing process and uses a consistent mask prediction threshold of 0.5 across models (Prytula et al., 3 Aug 2025). It also specifies that comparisons use single-scale inference and models trained until full convergence (Prytula et al., 3 Aug 2025).

These details matter because Revvity-25 images are originally 1080 × 1080, whereas the reported benchmarks are obtained after standardized resizing and cropping. A plausible implication is that boundary fidelity is evaluated under a scale-normalized inference regime rather than at native resolution, which may affect how methods exploit the high polygon detail present in the annotations.

5. Performance results and baseline status

The Revvity-25 benchmark is used to compare IAUNet against convolution-based, transformer-based, and query-based baselines, including Mask R-CNN, PointRend, Mask2Former, and MaskDINO (Prytula et al., 3 Aug 2025). The paper’s caption summarizes the outcome: “IAUNet outperforms strong query-based Mask2Former and MaskDINO baselines as well as other state-of-the-art models when training with fewer parameters. … IAUNet also efficiently scales with more queries while remaining efficient” (Prytula et al., 3 Aug 2025).

Selected reported results on Revvity-25 are as follows:

Model / backbone	AP	AP$_{50}$	AP$_{75}$
Mask R-CNN / R50	39.7	77.2	37.4
Mask2Former / R50	46.4	79.8	49.9
MaskDINO / R50	45.6	80.4	48.2
IAUNet / R50	49.7	82.1	54.8
Mask2Former / Swin-B	52.0	83.6	58.4
IAUNet / Swin-B	53.5	86.1	59.4
IAUNet / Swin-B, 300 queries	53.7	86.5	59.4

The full table also reports AP$_S$, AP$_M$, AP$_L$, parameter counts, and FLOPs (Prytula et al., 3 Aug 2025). In the IAUNet paper, the strongest reported Revvity-25 result is AP = 53.7 with Swin-B and 300 queries (Prytula et al., 3 Aug 2025).

The accompanying textual interpretation is that “IAUNet consistently outperforms other state-of-the-art models” on image-wise AP visualization and “visibly offers more detailed segmentation, capturing longer pixel relationships and effectively handling overlapping regions in some cases” (Prytula et al., 3 Aug 2025). Within the article’s scope, Revvity-25 therefore functions not only as a dataset release but also as a stress test for architectures designed to separate overlapping brightfield cell instances.

6. Relation to other cell-segmentation datasets

The IAUNet paper places Revvity-25 alongside LIVECell, EVICAN2, and ISBI2014 (Prytula et al., 3 Aug 2025). In that comparison, Revvity-25 is distinguished by three features repeatedly emphasized in the text: it is brightfield, it targets full-cell cytoplasm rather than only nuclei, and it prioritizes detailed annotations for cell borders and overlaps (Prytula et al., 3 Aug 2025).

The paper’s comparison can be summarized in conceptual terms. LIVECell is associated with phase-contrast imaging and very large scale; EVICAN2 spans brightfield + phase contrast + fluorescence; ISBI2014 is a more specialized cytology benchmark (Prytula et al., 3 Aug 2025). Revvity-25 is presented instead as a benchmark specifically for brightfield cancer cell images with high-detail cytoplasm contours and overlap-aware labeling (Prytula et al., 3 Aug 2025).

The paper further claims: “To our knowledge, this is the first dataset with accurate and detailed annotations for cell borders and overlaps” (Prytula et al., 3 Aug 2025). This is a strong positioning statement, but it should be understood exactly as given in the source. The paper does not claim that no prior whole-cell or overlap-rich datasets exist; rather, it claims novelty in the combination of accurate and detailed annotations for cell borders and overlaps in this setting (Prytula et al., 3 Aug 2025).

A plausible implication is that Revvity-25 is particularly relevant for evaluating modal and amodal segmentation formulations. The paper explicitly mentions this broader benchmarking role, even though it does not describe separate amodal labels or occlusion masks (Prytula et al., 3 Aug 2025).

7. Access, usage considerations, and interpretive boundaries

The dataset is reported as available via the IAUNet project repository: “Dataset available at: https://github.com/SlavkoPrytula/IAUNet” (Prytula et al., 3 Aug 2025). The same repository is also the stated code release location for IAUNet (Prytula et al., 3 Aug 2025). No DOI or dedicated standalone data portal is provided in the paper.

The paper does not specify license terms, directory structure, exact file naming conventions, or the on-disk annotation schema (Prytula et al., 3 Aug 2025). It also does not detail acquisition metadata such as microscope model or bit depth. Consequently, some practical aspects of reuse must be checked directly in the repository rather than inferred from the article text.

From the evidence given, Revvity-25 is best understood as a single-class brightfield whole-cell instance segmentation benchmark with the following defining characteristics: 110 high-resolution 1080 × 1080 images, 2937 expert-validated cell instances, 55/55 train-test split, and high-detail polygon annotations averaging 60 points per cell and reaching 400 points for complex structures (Prytula et al., 3 Aug 2025). Its principal research significance lies in the explicit treatment of overlapping cell cytoplasm, the precision of its contour annotation, and its use as a benchmark where strong modern instance segmentation baselines can be compared under a standardized protocol (Prytula et al., 3 Aug 2025).