Research map / Image generation / Super-resolution
SR-CACO-2: A Dataset for Confocal Fluorescence Microscopy Image Super-Resolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces SR-CACO-2, a public confocal fluorescence microscopy dataset of real LR/HR pairs (scales X2, X4, X8) from Caco-2 cells, and benchmarks 16 SISR methods on it.
Problem
Single-image super-resolution for confocal fluorescence microscopy lacks public datasets with real (not synthetic bicubic) low-resolution and high-resolution image pairs, and it is unclear how well SISR methods handle real microscopy degradation.
Approach
A new dataset of real paired confocal LR (/2, /4, /8) and HR tiles of Caco-2 3D cysts with three fluorescent markers, cropped into 512x512 ROI-only patches with registration, plus a benchmark of 16 SISR methods across four families and ROI-based evaluation.
Key results
- SRCNN reaches PSNR 35.08 (X2), 33.71 (X4) and 29.19 (X8) on ROI averages; post-upsampling and pre-upsampling methods are close, with VDSR and ENLCN best on some scales.
- Bicubic interpolation is outperformed by all benchmarked SISR methods on ROI-only evaluation, but methods still produce smooth outputs lacking accurate HR texture.
- Real LR images differ substantially from bicubic-downsampled LR (noisier, different intensity span), so bicubic-based training is not a substitute for real LR data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply SISR instead of bicubic interpolation | SISR methods on SR-CACO-2 · SR-CACO-2 test, ROI, X2 | PSNR: 32.52 vs 35.32 | Helped |
| Apply SISR instead of bicubic interpolation | SISR methods on SR-CACO-2 · SR-CACO-2 test, ROI, X4 | PSNR: 32.86 vs 34.12 | Helped |
| Apply SISR instead of bicubic interpolation | SISR methods on SR-CACO-2 · SR-CACO-2 test, ROI, X8 | PSNR: 28.49 vs 29.27 | Helped |
| Use L2 + λ SSIM instead of L2 loss only (ablation) | SRCNN · SR-CACO-2 test, CELL2, X2, ROI | PSNR: 33.13 vs 33.42 | Helped |
| Use L2 + λ SSIM instead of L1 loss only (ablation) | SRCNN · SR-CACO-2 test, CELL2, X2, ROI | PSNR: 33.03 vs 33.42 | Helped |
| Use L2 + λ SSIM instead of SSIM loss only (ablation) | SRCNN · SR-CACO-2 test, CELL2, X2, ROI | PSNR: 33.19 vs 33.42 | Helped |
| Use L2 + λ SSIM instead of L1 + λ SSIM (ablation) | SRCNN · SR-CACO-2 test, CELL2, X2, ROI | PSNR: 33.38 vs 33.42 | No clear effect |
Surprising findings
- Bicubic-generated LR images look smoother and preserve cell structure better than real microscope LR images, so they cannot substitute for real LR data.
- Some methods (e.g., SwinIR, GRL) collapse on certain cell channels at X2 on ROI evaluation (PSNR around 24-33), indicating training instability.
Limitations
- Only fixed (dead) Caco-2 cells are imaged, so live-cell dynamics and phototoxicity-relevant conditions are not captured.
- Only three markers from a single cell line; limited biological diversity of protein markers and cell types.
- Limited dataset diversity: no over/under-saturated or dim live-imaging conditions.
Experimental setup
Datasets: SR-CACO-2
Benchmarks: SR-CACO-2
Metrics: PSNR NRMSE SSIM
Baselines: Bicubic SRCNN VDSR DRRN MemNet NLSN DFCAN SwinIR EDSR (LIIF) ENLCN GRL ACT
Builds on: SRCNN VDSR EDSR SwinIR LIIF
Closest papers by meaning
- scSplit: Bringing Severity Cognizance to Image Decomposition in Fluorescence Microscopy (NeurIPS 2025 · similarity 0.86)
- Volume Tells: Dual Cycle-Consistent Diffusion for 3D Fluorescence Microscopy De-noising and Super-Resolution (CVPR 2025 · similarity 0.85)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.84)
- Micro-Bench: A Microscopy Benchmark for Vision-Language Understanding (NeurIPS 2024 · similarity 0.83)
- StructSR: Refuse Spurious Details in Real-World Image Super-Resolution (AAAI 2025 · similarity 0.83)
- Efficient Online Training for Zero-Shot Time-Lapse Microscopy Denoising and Super-Resolution (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.