Retention-Constrained Post-Training Quantization of Cellpose-SAM for Stem Cell Microscopy
Quantization study shows W8A16 and mixed W4/W8 preserve Cellpose-SAM segmentation on stem-cell microscopy, while ternary quantization fails on 169 of 176 fields.
The paper evaluates post-training quantization of Cellpose-SAM for iPSC microscopy deployment under a pre-specified retention criterion (95% cluster-bootstrap interval of F1 change above -0.02 per modality). On a 176-field panel spanning BBBC038 nuclei, BBBC039 U2OS fluorescence, and NIST iPSC images, weight-only W8A16 preserved instance F1 across all modalities. A sensitivity-guided mixed W4/W8 scheme with four INT8 exceptions achieved 6.76x weight compression with zero catastrophic failures, while ternary quantization reached 12.08x compression but failed on 169/176 fields. The authors propose modality-stratified retention auditing as the standard for compressed foundation models in regulated imaging.
- W8A16 quantization preserves Cellpose-SAM instance F1 across all tested imaging modalities
- Mixed W4/W8 scheme achieves 6.76x compression with 0/176 catastrophic failures
- Ternary quantization fails catastrophically on 169/176 fields despite 12.08x compression
- Proposes modality-stratified retention testing as an audit protocol for regulated deployment
Full article161 words · extracted from huggingface.co · click to collapse
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.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.21038