FluoCLIP: Stain-Aware Focus Quality Assessment in Fluorescence Microscopy

Conference: CVPR 2026 arXiv: 2602.23791 Code: To be confirmed Area: Multimodal VLM Keywords: Fluorescence microscopy, focus quality assessment, CLIP, ordinal regression, stain awareness

TL;DR

This paper proposes FluoCLIP, a two-stage vision-language framework that first performs stain-grounding to enable CLIP to learn the semantics of fluorescence stains, then conducts stain-guided ranking for stain-aware focus quality assessment. The paper also introduces FluoMix, the first multi-stain tissue-level fluorescence microscopy dataset for FQA.

Background & Motivation

Background: Focus quality assessment (FQA) is critical in microscopy imaging. Existing FQA methods are primarily designed for brightfield microscopy and rely on low-level features such as edges and gradients.

Limitations of Prior Work: In fluorescence microscopy, different fluorescent dyes exhibit distinct emission characteristics, signal-to-noise ratios, and background fluorescence, causing defocus degradation to manifest in a strongly stain-dependent manner. Simple edge-detection models (e.g., FocusLiteNN) perform well on brightfield data but are unstable on fluorescence data.

Key Challenge: Existing datasets do not capture stain-dependent variation in fluorescence microscopy — FocusPath is brightfield-based, while BBBC006 contains only two stains from in vitro cell lines.

Goal: (a) Construct a fluorescence FQA dataset spanning multiple tissues and stains; (b) enable FQA models to be aware of stain type and adjust focus predictions accordingly.

Key Insight: Focus quality in fluorescence images depends simultaneously on spatial sharpness and the spectral/semantic properties of stains. Visual features alone are insufficient; textual descriptions can provide complementary stain semantic information.

Core Idea: A two-stage CLIP adaptation strategy that first learns stain semantics and then performs stain-conditioned ordinal ranking.

Method

Overall Architecture

Stage 1 (Stain-Grounding): Learnable stain tokens and a lightweight adapter are appended to the CLIP text encoder, aligning stain textual representations with visual features via contrastive learning. Stage 2 (Stain-Guided Ranking): The stain embeddings learned in Stage 1 are used to condition the ranking prompts, enabling stain-aware FQA predictions.

Key Designs

1. Stain-Grounding Phase:

   • Function: Enables CLIP to understand the semantics of fluorescence stains (e.g., DAPI, Alexa-488), which lack meaningful correspondences in CLIP's original vocabulary.
   • Mechanism: Introduces learnable stain embeddings \(\mathbf{S}_l\) concatenated with context tokens to form pseudo-sentences. A lightweight adapter (single-layer self-attention + 2-layer MLP) enables the text encoder to acquire stain semantics, while the pretrained encoder is frozen to prevent semantic drift.
   • Design Motivation: Naively inserting stain names into CLIP prompts degrades performance due to the absence of semantic correspondences.

2. Stain-Guided Ranking Phase:

   • Function: Conditions focus level prediction on stain identity.
   • Mechanism: Learns base ranking embeddings \(\mathbf{R}^{base}\), which are combined with stain embeddings via a conditioning network \(f_\theta\) to produce stain-specific ranking embeddings \(\mathbf{R}^l_{k'}\). Interpolation is then used to generate intermediate-level ranking embeddings.
   • Design Motivation: The focus-appearance relationship differs across stains; a single shared ranking space cannot capture this heterogeneity.

3. FluoMix Dataset:

   • Function: Provides the first stain-aware FQA dataset.
   • Mechanism: Covers brain, lung, and liver tissues with up to four distinct stains per sample. Each field of view contains a 32-layer z-stack spanning the full range from sharp to severely blurred.
   • Design Motivation: Addresses the absence of multi-stain, multi-tissue FQA datasets for fluorescence microscopy.

Loss & Training

\(\mathcal{L}_{total} = \alpha \cdot \mathcal{L}_{CE} + \beta \cdot \mathcal{L}_{KL}\), where cross-entropy ensures classification alignment and KL divergence enforces ordinal consistency.

Key Experimental Results

Main Results (FluoMix, ResNet50 encoder)

Method	Accuracy (%)	PLCC ↑	SRCC ↑	MAE ↓
FocusLiteNN	-	0.621	0.624	1.610
CE (Cross-Entropy)	54.59	0.952	0.957	0.510
OrdinalCLIP	83.12	0.989	0.988	0.172
FluoCLIP	Best	Best	Best	Best

Stain Dependency Analysis

Dataset	SRCC (SF vs. Focus Level)	Inter-Stain Variation
FocusPath (brightfield)	-0.840 ± 0.092	Low (stain-independent)
BBBC006 (fluorescence)	-0.343 ± 0.292	High
FluoMix (fluorescence)	-0.528 ± 0.094	High

Key Findings

• Spatial frequency correlates strongly and stain-independently with focus level in brightfield data, but this correlation degrades significantly in fluorescence data with strong stain dependence.
• Directly inserting stain names into CLIP prompts not only fails to help but actually degrades performance, confirming the existence of a domain gap.
• In the two-stage design, the stain embeddings learned during the stain-grounding phase cluster with corresponding fluorescence images in the feature space.

Highlights & Insights

• Valuable task formalization: This is the first work to explicitly formulate FQA as a stain-aware ordinal regression problem, laying the groundwork for FQA in fluorescence microscopy.
• Elegant two-stage decoupled design: By addressing "which stain" before "which focus level," the method avoids entanglement between stain semantics and focus variation.
• The cross-domain CLIP adaptation strategy (frozen encoder + learnable tokens + lightweight adapter) is transferable to other domain-specific ordinal regression tasks.

Limitations & Future Work

• The scale and stain diversity of FluoMix remain limited; generalization to a broader range of fluorescent markers requires further validation.
• Only ResNet50 is used as the visual encoder; stronger ViT-based encoders may yield further improvements.
• Annotations rely on expert selection of the best-focus layer, which may introduce subjectivity-induced noise.
• The two-stage training pipeline increases workflow complexity.

Related Work & Insights

• vs. OrdinalCLIP: OrdinalCLIP is stain-agnostic; FluoCLIP achieves stain adaptation through stain-conditioned ranking embeddings.
• vs. NumCLIP: NumCLIP decouples numerical semantics, while FluoCLIP decouples stain semantics — the underlying idea is analogous but targets different domains.
• The multi-stage CLIP adaptation paradigm generalizes to other vision tasks requiring domain-specific concept grounding.

Rating

• Novelty: ⭐⭐⭐⭐ First formalization of stain-aware FQA task and dataset
• Experimental Thoroughness: ⭐⭐⭐ Experiments are primarily concentrated on a single dataset; cross-domain generalization experiments are limited
• Writing Quality: ⭐⭐⭐⭐ Task motivation is thoroughly analyzed; quantitative validation of stain dependency is convincing
• Value: ⭐⭐⭐⭐ Significant value to the biomedical image analysis community

Related Papers

• [ICLR 2026] VisJudge-Bench: Aesthetics and Quality Assessment of Visualizations
• [CVPR 2026] BriMA: Bridged Modality Adaptation for Multi-Modal Continual Action Quality Assessment
• [ICLR 2026] Grounding-IQA: Grounding Multimodal Language Models for Image Quality Assessment
• [CVPR 2026] A3: Towards Advertising Aesthetic Assessment
• [CVPR 2026] LFPC: Learning to Focus and Precise Cropping for MLLMs