Retention-Constrained Post-Training Quantization of Cellpose-SAM for Stem Cell Microscopy
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.
Community
Cellpose-SAM is used for segmenting cells and nuclei in microscopy images. Deploying it directly on laboratory instruments would enable analysis at the point of acquisition, but these devices have limited compute and memory compared with specialized workstations.
We therefore asked whether Cellpose-SAM could be compressed for local deployment without compromising segmentation accuracy, and whether compression affected imaging conditions differently. Cellpose-SAM compressed 12-fold degraded on 169 of 176 image fields, while a conservative mixed-precision scheme achieved a 6.76× reduction with no observed failures. These results show that compression must be evaluated across imaging conditions, not by model size or average accuracy alone.
The -0.02 bootstrap margin is the kind of threshold I can actually hand to an ops team — auditable, explainable, no vibes. What I'd want to see next is whether that margin holds when activations get quantized too, not just weights. Weight-only is the easy win; the moment you push INT8 activations through a real edge accelerator, the distribution shifts and your retention bound can quietly break. And is that interval per-modality or pooled across the stem cell types? If it's pooled, one noisy modality drags the whole threshold. I'd ship this with per-modality bounds and a re-quantization trigger.
Get this paper in your agent:
hf papers read 2609.21038 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 5
Browse 5 models citing this paperDatasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper