What it measures — PSNR quantifies pixel-level fidelity between a restored or processed image and a noise-free reference using the formula PSNR = 10·log₁₀(MAX²/MSE), where MAX is the maximum possible pixel value and MSE is the mean squared error across all pixels. Higher values (in dB) indicate less deviation from the reference. It is a direct proxy for pixel-level noise reduction and has no concept of structure, edges, or biological relevance.
Where it is used in microscopy — PSNR is the primary benchmark for fluorescence denoising algorithms. CARE (content-aware image restoration) established the convention of reporting PSNR on held-out test acquisitions to compare restoration fidelity across photon budgets; Noise2Void and its successors followed the same protocol. Super-resolution and deconvolution papers routinely pair PSNR with SSIM to characterize improvement relative to a diffraction-limited or widefield reference — it is practically ubiquitous in restoration benchmarks.
Where it breaks — Three failure modes matter for quantitative microscopy:
- The reference does not exist in practice. A true noise-free ground truth is not acquirable in live or low-light imaging — the photon-limited image is the only measurement. PSNR computed against a high-dose acquisition tests one acquisition condition against another, not fidelity to truth.
- MSE is intensity-weighted. Large, bright regions (cytoplasm, background) dominate the error term. Faint but biologically important structures — organelle membranes, sparse puncta, low-abundance probes — contribute negligibly to MSE. A model can maximize PSNR by blurring fine structure, which is exactly what Cell Painting profiling and segmentation rely on.
- It is decoupled from the readout. As image-quality-is-not-one-number argues, image-space metrics live at the lowest quality level; a restoration that raises PSNR can simultaneously erode profile reproducibility, segmentation accuracy, or transcript assignment.