henhao
Sign in
BioinformaticsbioRxiv
Preprint

ZEISS arivis Cloud: a cloud-based platform for deep learning model training and scalable bioimage analysis

Sreenivas Bhattiprolu, Manita Toor, Sebastian Soyer

Abstract

Modern biological imaging generates large, complex datasets that require scalable and reproducible image analysis methods. Deep learning has demonstrated strong performance on bioimage segmentation tasks, but training custom models has remained inaccessible to many researchers due to requirements for GPU infrastructure, programming expertise, and large annotated training datasets. ZEISS arivis Cloud is a browser-based platform for deep learning model training that addresses these barriers through partial annotation support, AI-assisted labeling with SAM (Segment Anything Model), pretrained model initialization, and automatically configured training pipelines requiring no machine learning expertise. The platform supports two segmentation tasks: semantic segmentation using a U-Net-style architecture with an EfficientNet encoder and PixelShuffle decoder, and instance segmentation based on Mask2Former with a Swin-Tiny backbone. Both pipelines incorporate microscopy-specific adaptations including smooth tiling, multi-channel input support, dataset-specific normalization, and partial-annotation-aware loss functions protected by patents US-20240078681-A1 and US-20250111519-A1. Trained models integrate directly with ZEISS arivis Pro for pipeline-based image analysis, ZEISS arivis Hub for parallel execution across large datasets, and ZEISS ZEN for content-aware guided acquisition. We describe the platform architecture, training methodology, segmentation architectures, reproducibility and versioning mechanisms, and FAIR compliance, and illustrate the complete workflow through two intestinal organoid imaging examples. arivis Cloud is freely accessible to student users; other users access the platform via subscription at https://www.arivis.cloud/.

PDFDOI
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
2
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (17 Aug 2026) · CC BY 4.0

Browser-based ZEISS platform trains U-Net/EfficientNet semantic and Mask2Former instance segmentation models from partial, SAM-assisted labels with fixed hyperparameters; described via two organoid demos, without accuracy benchmarks.

Training custom segmentation networks has usually required GPUs, programming skills and large annotated sets, and acquisition, segmentation and quantification tend to run as separate manual steps that are hard to reproduce. Compared with ilastik, QuPath, Cellpose and ZeroCostDL4Mic, the authors say what sets the platform apart is that browser-based training feeds directly into ZEN guided acquisition, arivis Pro pipelines and plate-scale arivis Hub runs.

Key results

  1. ClaimedTraining works with partial labels. Unlabeled pixels are left out of the loss: a weight map masks supervision in the semantic pipeline, and the Mask2Former pipeline uses a partial-annotation-aware loss. Users can start from roughly 20 annotated objects and add labels where predictions fail.

    The design is described and covered by patents US-20240078681-A1 and US-20250111519-A1. No experiment measures accuracy against the number of annotated objects or compares annotation effort with dense pixel labeling. · 4.3 Partial annotations; 4.7 Data-centric training strategy

  2. SuggestedAn organoid analysis ran end to end: a lumen model trained in arivis Cloud plus off-the-shelf Cellpose nuclei, combined in an arivis Pro measurement pipeline, then run as one arivis Hub batch that output a plate heatmap of nucleus counts.

    The pipeline was validated on 30 organoids per condition, then applied to 60 organoid datasets in a single Hub run (Figure 7). Cell number and Aldolase B expression are said to differ significantly between control and Wnt-inhibited organoids, but no test, effect size or p-value is reported. The lumen model is said to beat intensity thresholding, with no metrics. · 9.2 Scalable organoid analysis from model…; Figure legends

  3. SuggestedFor semantic segmentation, overlapping tile predictions are blended with weights that fall off toward tile edges. This removes the straight-edge artifacts seen when patches are simply stitched together.

    One qualitative cryo-electron microscopy example of mitochondria (Figure 5). Without blending, edge pixels were misclassified as mitochondria and unnaturally straight boundaries appeared; with smooth tiling no anomalies were visible. There is no quantitative comparison. · 5.3 Tiling strategies for large images; Figure legends

  4. ClaimedDetectors trained in arivis Cloud can drive ZEN guided acquisition: a low-magnification overview finds organoids, then the microscope takes high-resolution multichannel 3D stacks only at the detected positions.

    Murine Lgr5+ intestinal organoids in a 24-well plate on a Celldiscoverer 7: a widefield overview, then Airyscan 20x stacks of nuclei and E-cadherin (Figure 6). Thresholding is said to suffice for well-separated organoids. The text does not say which detector ran in the example shown, and neither the robustness gain nor the imaging-time savings are quantified. · 8.3 Intelligent acquisition in ZEN; 9.1 Content-aware guided acquisition in ZEN; Figure legends

  5. ClaimedEach training run is logged with a run identifier, linked dataset identifiers, JSON records of hyperparameters and normalization, and versioned model artifacts. Semantic models export as CZANN/CZMODEL files; instance models ship as containers through the ZEISS AI registry.

    This is a feature description only. The identifiers are internal to the platform rather than public persistent identifiers, and the authors note that exports are designed primarily for ZEISS software and not currently for deployment in arbitrary external environments. · 6.2 Reproducibility and versioning; Findable; Interoperable; 10. Discussion

Methods, data and reproducibility

A subscription browser platform on Microsoft Azure. Users label data with SAM suggestions and partial annotations, then train fixed-hyperparameter segmentation models, either semantic (U-Net-style, EfficientNet encoder, PixelShuffle decoder, 2D/3D) or instance (Mask2Former, Swin-Tiny), for use in ZEN, arivis Pro and arivis Hub. Two intestinal organoid workflows illustrate it.

  • Lgr5+ intestinal organoids in a 24-well plate; widefield overview plus Airyscan 20x 3D stacks (nuclei, E-cadherin) on a ZEISS Celldiscoverer 7 (guided acquisition example) · mouse (stated as murine) · n = 24-well plate
  • Intestinal organoids, control vs Wnt-inhibited, segmented for lumen, nuclei and cell bodies (scaled analysis example) · mouse (murine organoid per Figure 7 caption) · n = 30 organoids per condition (validation); 60 organoid datasets (Hub run)
  • FIB-SEM slice of a HeLa cell with partial mitochondria/background annotations (interface illustration) · human (HeLa cell line)
  • Cryo-electron microscopy image of a cell with mitochondria (smooth tiling illustration)
  • Phase contrast cell images (semantic vs instance segmentation illustration)
Methods & tools
ZEISS arivis Cloud, SAM (Segment Anything Model), U-Net-style network with EfficientNet encoder, Mask2Former (Swin-Tiny backbone), ImageNet pretrained weights, Cellpose, ZEISS arivis Pro, ZEISS arivis Hub, ZEISS ZEN, ZEISS Celldiscoverer 7 with Airyscan, Microsoft Azure
Reagents & environment
Runs server-side on Microsoft Azure; no software versions are given. Training configurations are stored as JSON. Semantic models export as CZANN/CZMODEL files; instance models as runnable containers through the ZEISS AI registry.
Reproducibility gaps
Subscription platform at https://www.arivis.cloud/ (free for students). No code, images or trained models released. Hyperparameters hidden; instance tile merging proprietary; partial-annotation loss patented; exports designed mainly for ZEISS software rather than arbitrary external environments.
Limitations (6)
  • Trained models and pipelines are designed primarily for the ZEISS ecosystem. CZANN/CZMODEL exports and instance-model containers are not currently meant to run in arbitrary external environments. · stated by the authors
  • There is no community model repository, so trained models cannot be shared publicly in a way that would let external readers of a study fully reproduce it. · stated by the authors
  • Training metrics (IoU, loss curves) are not shown prominently in the interface; users who need them for a publication must retrieve the stored artifacts. · stated by the authors
  • Instance segmentation trains on all annotations with no held-out validation split, and the 3D semantic mode starts from random weights. No pipeline uses microscopy-specific pretraining. · stated by the authors
  • No quantitative evaluation: no IoU or average precision values, no annotation-time measurements, and no comparison with ilastik, Cellpose, ZeroCostDL4Mic or thresholding. Claims of fewer labels, better lumen segmentation and shorter imaging time are untested here. · noted in this summary
  • Hidden hyperparameters, a proprietary instance tile-merging algorithm and a patented loss mean the training pipelines cannot be independently re-implemented or audited from this description. · noted in this summary

Who should read:Read this if you use ZEISS microscopes or arivis software and need to know how arivis Cloud trains, versions and deploys segmentation models (for example, to write a methods section); it contains no accuracy benchmarks.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISHaberl '18Gomez-de-Ma '21Chamier '21Przygodzki '26Consortium '24Rasse '20
71% semantic matchCell Biology
M. Haberl, C. Churas, L. Tindall+8 more·bioRxiv·Jun 21, 2018