What to Check When Compressing a Cell-Imaging Foundation Model ── Evaluating quantized Cellpose-SAM with a pre-specified retention criterion per imaging modality

What problem does the paper address?

Induced pluripotent stem cell (iPSC) culture increasingly uses large foundation models to segment cells and nuclei in microscope images. Cellpose-SAM^[3] is a leading example: it brings the backbone of a general image foundation model into a framework built for cell images.

The problem is where such a model runs. Culture rooms and quality control labs often have laboratory CPUs or small edge devices rather than GPU servers. Quantization, which replaces weights with lower-precision numbers, is the natural way to make the model lighter. In regulated imaging, however, being able to show that the lighter model still gives the same results matters as much as the savings.

The paper asks how that case should be made. Quantization studies usually report one overall accuracy number. A single number can hide a collapse under one imaging condition. The author sets out to make compression evaluation auditable.

What does it propose?

The core idea is to fix the pass criterion before evaluating any compression scheme. The criterion, called retention, is defined as follows. Compute the 95% cluster-bootstrap interval of the mean change from the FP32 model. That interval must stay above a fixed -0.02 margin for every imaging modality, with no exceptions.

The evaluation uses a stratified panel of 176 fields. It covers BBBC038 nuclei^[4], BBBC039 U2OS fluorescence images^[5] and NIST iPSC images, across different cell-density regimes. Against this panel the paper compares weight-only W8A16, a sensitivity-guided mixed W4/W8 scheme, and ternary weight-only quantization.

The weight of the contribution lies in the procedure rather than in a new method. Set the criterion in advance, stratify by modality, and decide on intervals. The structure is close to the non-inferiority logic used in clinical trials.

What did it show?

The results stated in the abstract are limited to the following. Weight-only W8A16 preserved instance F1 across all modalities. The mixed W4/W8 scheme, which keeps four layers as INT8 exceptions, cut weight storage by a factor of 6.76. It showed no observed catastrophic failures (0/176 fields) and matched W8A16 at this sample size.

By contrast, ternary weight-only quantization reached 12.08x compression but failed catastrophically on 169 of 176 fields. From this the author argues that compression should be judged by modality-stratified downstream retention, not by a single accuracy figure, and claims to provide a reproducible protocol for auditing compressed foundation models in regulated stem-cell imaging.

This paper is a preprint that has not been peer reviewed. The results above are the author's claims.

A worked example

Consider a group that uses a segmentation model to count cells and assess morphology during culture quality checks. It wants to run the model on bench-side devices and is looking at compression.

The second approach takes more work, but it lets the team explain later what criterion was used, what was checked, and why the model was accepted. That is what the paper means by auditable.

What is not new, and where are the limits?

Quantization, weight-only quantization, and mixed precision guided by layer sensitivity are all existing techniques. Setting a margin in advance and deciding on an interval is familiar from non-inferiority testing. The contribution is to bring these into the evaluation of a compressed cell-imaging foundation model and to assemble them into a protocol.

The limits are clear. First, the evaluation covers 176 fields, and the abstract itself qualifies its claim with "at this sample size". Zero observed failures for the mixed scheme does not mean failures cannot occur. Second, the abstract does not explain why the margin was set at -0.02. The right margin depends on the use, and whether it is acceptable for quality decisions needs separate work.

Third, what was measured is storage reduction and F1 retention. The abstract does not report how inference speed or memory use on a CPU actually changed. Fourth, the paper has a single author and covers one model and three image sources. Whether the conclusions carry over to other segmentation models or other cell types has not been tested.

Why is this topic clustering now?

In this collection run the paper formed a cluster of two (cluster size 2) with a paper on quantizing medical LLMs^[2]. That paper's title, "When Quantization Preserves Accuracy but Not Evidence", points to explanation-aware post-training quantization. Shared keywords include quantization, post-training, medical, evidence and preserves.

What links the two is the view that in medicine, "average accuracy was preserved" is not enough to accept a compressed model. One asks about retention under each imaging condition; the other asks about retention of the evidence behind a decision. The modalities differ, image versus language, but both try to state explicitly what must be preserved.

Attention is modest: at collection time the paper had 2 upvotes on Hugging Face. Upvotes indicate visibility, not correctness. A cluster of size 2 is best read as a topic that is only starting to form.

Where does it connect to pharma and regulatory work?

In manufacturing regenerative and cell therapy products, image analysis can feed into process control and release decisions. If an AI model does that analysis, compressing the model is itself a change to a validated system.

The paper's protocol fits change-control thinking well. Document acceptance criteria in advance, verify under each condition, and record the results. That is the flow expected in computerized system validation under GxP. Being able to write "met the pre-specified criterion under every imaging condition", rather than "average accuracy was the same", makes the case easier to explain in inspections and audits.

Still, this is a research-stage evaluation, not a demonstration that the protocol meets any regulator's standard. The margin, the representativeness of the evaluation images and the definition of failure all have to be set again by the user for their own process.

Summary

This preprint proposes that when a cell-imaging foundation model is compressed, the pass criterion should be fixed first and checked with intervals under each imaging condition. The author reports that moderate quantization retained performance while aggressive compression failed on most fields. Most of the techniques are established and the evaluation is small. It is still useful material for thinking about what to record when an AI model changes in a regulated setting.