A Unified Resolution-Conditioned Framework for Orthogonal Line-Scanning Image Fusion

Technical Summary: A Unified Resolution-Conditioned Framework for Orthogonal Line-Scanning Image Fusion

1. Problem Statement

Laser line-scanning microscopy (LLM) enables rapid volumetric imaging by replacing point illumination with line illumination, achieving 10–100× speedups over traditional point-scanning methods. However, this parallelization introduces severe anisotropic spatial resolution: the system resolves features sharply along the scanning axis but exhibits significant blurring along the orthogonal axis due to the confocal slit. The degree of anisotropy is quantified by the resolving-power ratio $\rho = \Delta_{high} / \Delta_{low}$, which can drop to values as low as 0.014 in practical settings.

While acquiring two orthogonal line scans (X and Y) and fusing them computationally can recover near-isotropic resolution, existing deep learning approaches face critical limitations:

• Configuration Specificity: Current state-of-the-art models (e.g., Attn-DualUnet) require training a separate network for every specific optical configuration (slit width and numerical aperture).
• Multi-Configuration Failure: Training a single model on mixed configurations without explicit conditioning leads to catastrophic performance collapse (dropping from ~38 dB to ~24 dB), as the network cannot disambiguate between fundamentally different degradation structures.
• Classical Limitations: Traditional Fourier-domain methods fail when sampling is coarse, a regime where LLM is most needed.

2. Methodology

The authors propose CondLAformer, a unified, resolution-conditioned framework that processes arbitrary slit configurations using a single trained model. The architecture is built upon a U-Net topology with four encoder-decoder levels, incorporating three core innovations grounded in the physical analysis of the Point Spread Function (PSF).

2.1. Physics-Grounded Data Generation
The authors derived a separable PSF approximation where the line PSF's vertical extent scales linearly with the slit width ($FWHM \approx \alpha \cdot s$). This allows for the efficient generation of training pairs at arbitrary slit values via 1-D convolutions, verified against measured optical data with 48.3 dB accuracy. A multi-resolution dataset was constructed spanning 15 slit configurations ($s \in [8, 52]$).

2.2. Resolution-Conditioned Architecture (FiLM)
To resolve the ambiguity between different degradation levels, the network receives the resolution ratio $\rho$ as a continuous conditioning signal.

• Feature-wise Linear Modulation (FiLM): The normalized slit parameter is processed by a shared MLP to generate scale ($\gamma$) and shift ($\beta$) parameters. These are injected into every encoder/decoder block after LayerNorm. This allows the network to dynamically adjust feature operating points based on the specific degradation severity, enabling a single model to handle both mild and severe anisotropy.

2.3. Adaptive Rank Enhanced Linear Attention (Adaptive RELA)
Standard Linear Attention suffers from low-rank constraints, while standard Rank Enhanced Linear Attention (RELA) uses a fixed kernel size that mismatches the varying blur scales of different slit widths. The authors introduce Adaptive RELA with two specific extensions:

• Multi-Scale Depthwise Convolutions: Instead of a fixed $5\times5$ convolution for rank enhancement, the model employs three parallel depthwise convolutions with different receptive fields (3px, 13px, 25px). Their blending weights are learned as a function of the resolution ratio $\rho$, allowing the network to select the appropriate local processing scale.
• Learnable Attention Temperature: A learnable temperature parameter $\tau(r)$ modulates the linear attention distribution. Under severe degradation (small slits), shallow layers increase $\tau$ to sharpen attention on sparse informative tokens; deeper layers decrease $\tau$ to enable broader semantic aggregation. This mechanism provides degradation-proportional selectivity.

2.4. Network Components
The core block is a Conditional Dual-Attention Block, which integrates:

1. Adaptive RELA: For global spatial modeling with ratio-adaptive selectivity.
2. Channel Attention Block (CAB): For channel-wise recalibration.
3. Convolutional Gated Feed-Forward Network (CG-FFN): For local feature refinement using gated depthwise convolutions.

3. Key Contributions

1. Unified Multi-Configuration Framework: The first deep learning approach to fuse orthogonal line scans across a continuous range of slit configurations using a single model, eliminating the need for per-configuration training.
2. Adaptive RELA Mechanism: A novel extension of linear attention that dynamically adjusts both the local receptive field (via multi-scale convolution blending) and the global attention selectivity (via learned temperature) based on the degradation parameter.
3. Physics-Informed Dataset: A dense training set generated via separable PSF simulation, covering the full practical range of slit widths and validated against measured optical data.
4. Systematic Evaluation: A comprehensive benchmark demonstrating that the unified model outperforms both classical methods and per-configuration specialists, particularly in generalizing to unseen intermediate configurations.

4. Experimental Results

The model was evaluated on 281 real-structure biological images (microtubules, F-actin, etc.) across 18 slit values (12 trained, 6 interpolated).

• Performance Gains: The full CondLAformer v3 (Adaptive RELA + FiLM) achieved 34–40 dB PSNR across all configurations.
  • It significantly outperformed the unconditioned multi-slit training (which collapsed to 24.3 dB).
  • It surpassed the standard RELA + FiLM variant (CondLAformer v2) by ~2 dB in the small-slit regime ($\rho \geq 0.05$), where anisotropy is subtle and directional signals are weak.
• Generalization: The model demonstrated smooth generalization to unseen interpolated configurations (e.g., slit widths not present in training) with no visible artifacts or performance discontinuities.
• Comparison to Specialists: While per-slit specialist models achieved high performance on their specific training slit, they degraded by 4–9 dB when applied to other configurations. In contrast, the unified CondLAformer maintained consistent high performance (>35 dB) across the entire continuum.
• Ablation Study:
  • FiLM Conditioning: Provided the largest gain (+11.5 dB over unconditioned, +7.3 dB over unconditioned multi-slit), proving essential for resolving configuration ambiguity.
  • Adaptive RELA: Provided consistent gains, specifically critical for handling the nuanced degradation of small slits where fixed attention fails.

5. Significance and Claims

The paper claims that the proposed framework addresses a fundamental practical bottleneck in line-scanning microscopy: the inability to deploy a single fusion model across varying acquisition parameters.

• Physical Prior Encoding: The authors argue that the success of the architecture stems from its alignment with optical physics. The use of directional attention encodes the physical prior that each scan is trustworthy only along its resolved axis.
• Separation of Concerns: Analysis of learned behaviors reveals a principled division of labor: multi-scale convolutions provide fixed, depth-appropriate local biases; FiLM acts as a "soft bypass" gating the network between fusion regimes; and the attention mechanism (modulated by temperature) implements the actual adaptive receptive field.
• Practical Deployment: The framework enables microscopy facilities with variable slit settings to utilize a single model for the entire continuum of configurations, including intermediate values never seen during training, thereby removing the overhead of training and maintaining dozens of specialized models.

The authors conclude that this paradigm of physics-informed conditioning of flexible attention mechanisms offers a scalable solution for imaging modalities with continuously variable acquisition parameters.