Retention-Constrained Post-Training Quantization of Cellpose-SAM for Stem Cell Microscopy
Sebastián A. Cruz Romero
Capicú Technologies
Published September 17, 2026 · on the feed September 21, 2026 · ▲ 3 · arXiv 2609.21038
The abstract
Induced pluripotent stem cell (iPSC) culture increasingly relies on segmentation foundation models, yet deployment on laboratory CPUs and edge hardware requires compression schemes that are both efficient and auditable. We present a deployment-oriented evaluation of compressed Cellpose-SAM using a pre-specified retention criterion: the 95% cluster-bootstrap interval of mean change from FP32 must remain above a fixed -0.02 margin for every imaging modality. On a stratified 176-field panel spanning BBBC038 nuclei, BBBC039 U2OS fluorescence, and NIST iPSC images across density regimes, weight-only W8A16 preserves instance F1 across all modalities. A sensitivity-guided mixed W4/W8 scheme, using four INT8 exceptions, achieves a 6.76x reduction in weight storage with no observed catastrophic failures (0/176 fields), matching W8A16 at this sample size. In contrast, ternary weight-only quantization achieves 12.08x compression but fails catastrophically on 169/176 fields. These results demonstrate that compression should be evaluated by modality-stratified downstream retention rather than single-number accuracy, and establish a reproducible protocol for auditing compressed foundation models in regulated stem-cell imaging.
The Brief
2:13A short podcast summary of the paper. Written and voiced by AI.
Transcript
Retention-Constrained Post-Training Quantization of Cellpose-SAM for Stem Cell Microscopy. By Sebastián A. Cruz Romero, from Capicú Technologies. Published on the seventeenth of September. No code repository is linked to it.
The central claim concerns how compressed models ought to be judged. Compression should be evaluated by downstream retention, stratified by imaging modality, rather than by a single accuracy number. The setting is induced pluripotent stem cell culture, which increasingly relies on segmentation foundation models, and where deployment on laboratory CPUs and edge hardware needs compression that is efficient and also auditable.
The model is Cellpose-SAM, and the retention criterion is fixed in advance. For every imaging modality, the ninety-five percent cluster-bootstrap interval of the mean change from full precision must stay above a margin of minus zero point zero two. The test panel is stratified, with one hundred and seventy-six fields spanning BBBC038 nuclei, BBBC039 U2OS fluorescence, and NIST iPSC images across density regimes.
Weight-only quantization to eight bits, with sixteen-bit activations, preserves instance F1 across all modalities. A mixed four-bit and eight-bit scheme, guided by sensitivity and keeping four INT8 exceptions, cuts weight storage by six point seven six times. It shows no catastrophic failures, none of the one hundred seventy-six fields, which matches the eight-bit scheme at this sample size.
Ternary weight-only quantization compresses further, by twelve point zero eight times, but fails catastrophically on one hundred sixty-nine of the one hundred seventy-six fields.
These results come from one model and one panel, and the mixed scheme's clean record is stated at this sample size, not beyond it. What the author establishes is a reproducible protocol for auditing compressed foundation models in regulated stem-cell imaging. The contrast between schemes is the substance of the claim. A scheme that looks attractive on compression alone, the ternary one, is precisely the one that the stratified, pre-specified test rejects.
Measured this way, compression becomes something a regulated lab can audit: a criterion set before testing and checked modality by modality, rather than one number averaged over everything.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A big model, a small machine, and a lab that has to trust the result.
Your stem-cell lab segments microscope images with a large foundation model, but the lab machines are CPUs and edge boxes with no GPU. The model has to be compressed before it can run there.
The road ahead
- 1 · The problem ~6 min A big cell-segmentation model must be shrunk to run on lab CPUs, and one overall accuracy score may hide a failure on one imaging type.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Judge each compressed model by how much of the full-precision model's cell-finding it keeps, against a bar fixed in advance, on each imaging type separately.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Set the bar, compress the model several ways, test each version on a stratified panel, and accept only versions that clear the bar on every modality.
- 6 · The experiment ~6 min You’ll guess: Ternary weight-only quantization shrinks Cellpose-SAM's weights the most. On how many of the 176 test fields did it fail catastrophically?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.