Research map / Image generation / Image restoration
scSplit: Bringing Severity Cognizance to Image Decomposition in Fluorescence Microscopy
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.
Proposes scSplit (indiSplit), an InDI-based iterative-restoration method for fluorescence image unmixing that predicts the mixing ratio with a regressor and uses severity-cognizant input normalization to handle unknown superposition severity.
Problem
Image decomposition of superimposed fluorescence microscopy channels (and bleedthrough removal) when the relative mixing ratio of the structures is unknown and varies across inputs, which existing methods trained at a fixed ratio do not handle.
Approach
Combines the InDI linear-mixing training framework with a regressor that estimates the mixing ratio t, a severity-cognizant input normalization (SCIN) that makes per-input statistics independent of t, and session-level aggregation of t estimates, so one network handles all mixing ratios for both unmixing and bleedthrough removal.
Key results
- indiSplit outperforms InDI by about 1.2 dB PSNR on average across regimes and datasets through the SCIN normalization alone (indiSplit-AbONE vs. InDI).
- indiSplit beats the best muSplit variant augmentation by about 2.4 dB PSNR on average.
- On HTT24 with intensity-altered real superimposed inputs, indiSplit consistently outperforms U-Net and denoiSplit by a substantial PSNR margin.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SCIN instead of two-image mean-std normalization (ablation) | scSplit · Four microscopy datasets, all input regimes | PSNR: +1.2db PSNR on average | Helped |
| add SCIN instead of two-image mean-std normalization (ablation) | scSplit · PaviaATN, Hagen, HTT24, and BioSR; all input regimes | PSNR | Helped |
| add aggregation of estimated t across images in a session (ablation) | scSplit · All reported microscopy tasks | PSNR | Helped |
| add the Reg network to estimate t instead of fixing t=0.5 (ablation) | scSplit · Several microscopy tasks, especially Dominant and Weak regimes | PSNR | Helped |
| use mixing-ratio augmentations during training (ablation) | Deep-LC · Microscopy unmixing at mixing ratios away from w=0.5 and at w=0.5 | PSNR: 2.4db PSNR | Mixed |
| use SCIN rather than first-layer instance normalization (ablation) | scSplit · Three input regimes across two datasets | PSNR: 0.33 dB | Helped |
| vary the assumed inference t (ablation) | scSplit · GoPro motion deblurring test | PSNR: 35.0; 37.3; 36.2; 32.4 | Mixed |
| use median, mode, WgtSum, or WgtProd instead of mean aggregation (ablation) | scSplit · BioSR, HTT24, and HTLIF24 | mixing-ratio estimation | No clear effect |
Surprising findings
- Under standard per-dataset mean/std normalization, the expected variance of superimposed patches depends on the mixing ratio t, so normalization cannot be fixed without knowing t, which existing unmixing methods overlook.
- Alternative aggregation schemes (median, mode, weighted sums) gave no significant advantage over simple mean aggregation of mixing-ratio estimates.
Limitations
- Not optimized for noisy data; requires a two-step denoising-then-unmixing pipeline.
- Limited to two channels; extending to more than two structures is future work.
- Assumes linear superposition, which may not hold for interference-based imaging modalities.
Experimental setup
Datasets: Hagen et al. BioSR HTT24 HTLIF24 PaviaATN
Benchmarks: Hagen et al. (Actin vs. Mitochondria) BioSR (ER vs. Microtubules) HTT24 (SOX2 vs. Golgi) HTLIF24 (Microtubules vs. Centromere) PaviaATN (Actin vs. Tubulin)
Metrics: PSNR SSIM LPIPS Dice dissimilarity (downstream segmentation)
Baselines: U-Net Lean-LC muSplit Regular-LC muSplit Deep-LC muSplit denoiSplit MicroSplit InDI
Builds on: InDI muSplit denoiSplit MicroSplit
Closest papers by meaning
- SR-CACO-2: A Dataset for Confocal Fluorescence Microscopy Image Super-Resolution (NeurIPS 2024 · similarity 0.86)
- FluoCLIP: Stain-Aware Focus Quality Assessment in Fluorescence Microscopy (CVPR 2026 · similarity 0.85)
- Efficient Online Training for Zero-Shot Time-Lapse Microscopy Denoising and Super-Resolution (AAAI 2025 · similarity 0.84)
- MorphoHELM: A Comprehensive Benchmark for Evaluating Representations for Microscopy-Based Morphology Assays (NeurIPS 2026 · similarity 0.84)
- ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy (ICML 2025 · similarity 0.84)
- CellCLIP - Learning Perturbation Effects in Cell Painting via Text-Guided Contrastive Learning (NeurIPS 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.