In an image-based spatial-omics experiment, a bright signal becomes useful only when its location is understood. Which cell contains it? Is it near the cell boundary? Could two adjacent cells have been treated as one? The segmentation step—the assignment of image pixels to individual cells—can affect the measurements used later in the analysis.
For AINA, the LabCobot, this suggests a research workflow connecting image records, a segmentation model and reviewable measurements. The VISTA-2D connection described here is a proposed integration. SupraWorx has generic camera-image intake and record-linking code; a microscopy-specific adapter and a validated VISTA-2D analysis path would still need to be developed.
What VISTA-2D contributes
NVIDIA describes VISTA-2D as a training and inference pipeline for cell segmentation, packaged in the MONAI bundle format. For a research evaluation, record the selected model and code version, supported inputs, preprocessing and output representation. Performance reported for the manufacturer's test datasets or hardware should not be substituted for measurements on the laboratory's own configuration.
The task also has established research alternatives. Stringer and colleagues' Cellpose study (2021) developed a segmentation method using varied cellular images. Greenwald and colleagues (2022) developed TissueNet and Mesmer for tissue-cell segmentation and demonstrated extraction of cellular features from multiplexed images. These papers provide methods and datasets to consider when choosing a comparison; neither evaluates an AINA integration.
Define the object before selecting the model
A nucleus, an entire cell and a colony are different targets. Specify which one the experiment needs, which image channels show its boundaries and how ambiguous examples will be annotated. A model that finds nuclei does not automatically provide whole-cell boundaries suitable for assigning every surrounding signal.
SupraWorx also contains a geometric image helper that detects circular cell fields or colonies and generates crops. That helper addresses a different task from learned tissue-cell segmentation. Its output should not be treated as evidence that a VISTA-2D workflow already exists.
Build a small, informative reference set
Select images that reflect the intended experiment: tissue preparation, staining, microscope settings, pixel size and expected cell density. Include difficult regions such as touching cells, weak boundaries and uneven illumination. Keep separate samples for tuning and final evaluation; neighbouring tiles from one specimen should not silently appear on both sides of the comparison.
Ask qualified annotators to follow a written boundary convention and record uncertain regions. Preserve those decisions with the reference masks. This makes it possible to distinguish disagreement about the biological object from an error made by the model.
Inspect errors that change the research result
Overlay predicted boundaries on the source image and review at least four error types: missed cells, spurious objects, two cells merged into one and one cell split into several. Report object counts and a defined boundary-overlap measure alongside these examples. A single average score can hide a recurring failure in a particular tissue region.
Then test a downstream measurement. For example, compare a marker-intensity estimate using the reference boundaries and the model's boundaries. If a merged region combines neighbouring cells, does it change the resulting classification or summary? The appropriate acceptance criterion should follow the experiment's question and be agreed before the final test.
Make the proposed connection reproducible
An AINA adapter should preserve the original microscopy file, channel names, pixel scale and specimen coordinates. If images are tiled, record how tile coordinates map back to the specimen. Keep model version, parameters, predicted masks, reviewed corrections and derived measurements together so another researcher can reconstruct a result.
Measure processing time and memory for the complete path, including reading files, preprocessing, segmentation and storing results. Check large images, failed inputs and interrupted jobs on the intended hardware. The useful outcome of an initial pilot is a documented comparison and a clear decision about the next experiment—not a claim of diagnostic capability or a general throughput figure.
