A Unified Resolution-Conditioned Framework for Orthogonal Line-Scanning Image Fusion
Unified fusion framework based on Rank-enhanced linear attention with FiLM conditioning achieves 34-40dB PSNR across multiple optical slit configurations.
Key Findings
Methodology
This work introduces a resolution-conditioned image fusion framework utilizing Rank Enhanced Linear Attention (RELA). FiLM modules provide continuous conditioning on the resolution ratio, enabling a single model to adapt across slit widths. Adaptive RELA replaces fixed kernels with multi-scale dilated depthwise convolutions, modulated by ratio-dependent attention temperature. Training data are generated via a physics-grounded separable PSF model, verified at 48.3dB accuracy, covering 15 slit configurations. The model demonstrates PSNR of 34-40dB across configurations, outperforming unconditioned and single-slit models, with smooth generalization to unseen intermediate configurations.
Key Results
- The proposed model achieves 34-40dB PSNR across 15 slit configurations, significantly surpassing unconditioned multi-slit training (24.3dB) and per-slit models (38dB). It generalizes well to intermediate unseen configurations without artifacts, confirming the effectiveness of the resolution conditioning. Ablation studies show FiLM resolves configuration ambiguity, global linear attention captures long-range directional correlations, and adaptive temperature boosts performance by 2dB in near-isotropic regimes.
- In experiments, the model maintains high fidelity in challenging anisotropic conditions, with PSNR improvements of over 10dB compared to classical methods like Fourier-domain weighted fusion. The adaptive multi-scale approach dynamically adjusts receptive fields, leading to robust performance across diverse slit widths. The learned attention temperature effectively sharpens focus under severe degradation, further enhancing reconstruction quality.
- Overall, the framework demonstrates strong cross-configuration generalization, enabling continuous interpolation between configurations and providing a practical solution for flexible, high-quality volumetric imaging in microscopy.
Significance
This research addresses a critical bottleneck in line-scanning microscopy—anisotropic resolution—by enabling a single model to adapt seamlessly across varying optical configurations. It bridges the gap between physics-based modeling and deep learning, offering a versatile, high-performance solution that can be deployed in real-world biological imaging. The approach enhances the capability of microscopes to produce near-isotropic images without hardware modifications, facilitating more accurate 3D reconstructions and dynamic imaging. Its generalization ability reduces the need for multiple models, streamlining workflows and reducing costs, thus impacting both academic research and clinical diagnostics.
Technical Contribution
The key technical innovation lies in integrating FiLM-based continuous conditioning with an adaptive RELA mechanism that employs multi-scale dilated depthwise convolutions and a learnable attention temperature. This design overcomes the low-rank limitation of standard linear attention, providing full-rank feature diversity across configurations. The architecture combines global linear attention with local detail capture via gated depthwise convolutions, enabling robust, multi-scale feature extraction. The physics-grounded data synthesis further ensures realistic training, while the end-to-end trainable framework achieves superior generalization and interpolation capabilities, setting new standards for multi-configuration image fusion.
Novelty
This work is the first to unify multi-configuration line-scanning image fusion within a single, resolution-conditioned transformer framework. Unlike prior methods that train separate models or naive multi-slit models, it employs FiLM conditioning and adaptive RELA to dynamically adjust receptive fields and attention sharpness based on slit width. The physics-informed data generation and continuous parameter conditioning enable smooth interpolation between configurations, representing a significant leap in the field of optical microscopy and deep learning-based image restoration.
Limitations
- The model's performance heavily depends on the accuracy of the physical PSF model; deviations in real optical systems may reduce effectiveness.
- Training relies on the linear relationship between slit width and PSF FWHM, which may not hold in complex or aberrated optical setups.
- High-resolution inference incurs substantial computational costs, necessitating further optimization for real-time applications.
Future Work
Future directions include extending the framework to handle more complex PSF models, incorporating aberration correction, and applying the approach to other imaging modalities such as multi-photon or super-resolution microscopy. Enhancing computational efficiency and exploring unsupervised or self-supervised training strategies could further improve practical deployment. Additionally, integrating adaptive hardware control with the software framework may enable real-time, configuration-aware imaging in diverse biological environments.
AI Executive Summary
Line-scanning microscopy offers rapid volumetric imaging but suffers from anisotropic resolution, limiting image quality and interpretability. Traditional fusion methods like Fourier-domain weighted fusion perform well only under high sampling conditions, failing in coarse or low-photon regimes. Deep learning approaches such as DualUnet and Attn-DualUnet introduced directional attention mechanisms but require separate models for each optical configuration, making them impractical for variable slit widths. This paper presents a unified, resolution-conditioned framework based on Rank Enhanced Linear Attention (RELA), augmented with FiLM modules for continuous parameter conditioning. The core innovation is Adaptive RELA, which employs multi-scale dilated depthwise convolutions and a learnable attention temperature, enabling the model to adapt seamlessly across slit widths. Using a physics-grounded PSF model, the authors generate a comprehensive training dataset spanning 15 slit configurations, verified at 48.3dB accuracy. Experimental results demonstrate that the proposed model achieves 34-40dB PSNR across configurations, outperforming unconditioned and per-slit models significantly. It generalizes smoothly to unseen intermediate configurations without artifacts, confirming its robustness and flexibility. Ablation studies highlight the importance of FiLM in resolving configuration ambiguity, the efficacy of global linear attention in capturing long-range directional relationships, and the benefit of adaptive temperature in challenging near-isotropic regimes. This work advances the field by enabling high-quality, multi-configuration image fusion with a single, adaptable model, promising broad impact in biological imaging and beyond.
Deep Dive
Plain Language Accessible to non-experts
想象你在厨房做饭，不同的锅子和火候会影响菜的味道。有时候用大火炒菜，菜会焦掉；用小火慢炖，味道更浓。这就像是用不同的光学设置拍照，得到的图片清晰度也不同。这个研究就像发明了一台智能炉子，可以根据你要做的菜自动调节火力，不管锅子大小或火力强弱，都能做出好吃的菜。它学习了不同火候的效果，然后用一种聪明的方法，把这些信息融合在一起，让你不用换炉子，也能做出完美的菜。这让厨房变得更方便，也能做出更好吃的饭。
ELI14 Explained like you're 14
想象你在玩一款超级酷的游戏，不同的关卡难度不同。有时候你会遇到特别难的关卡，需要用不同的策略才能过关。这个研究就像是设计了一个超级聪明的游戏助手，它可以根据关卡的难度自动调整策略，让你无论遇到多难的关卡都能顺利过关。它学习了各种关卡的特点，比如难度高或低，然后用一种特别的方法，把这些信息融合在一起，帮你做出最好的决策。这样一来，不管关卡怎么变，它都能帮你轻松应对，让游戏变得更有趣、更顺畅！
Abstract
Laser line-scanning microscopy enables fast volumetric imaging but produces anisotropic lateral resolution. Orthogonal line scans provide complementary directional information that can recover near-isotropic resolution, yet existing deep-learning methods require a separate model for each optical configuration. We present a unified, resolution-conditioned fusion framework based on Rank Enhanced Linear Attention (RELA). Feature-wise Linear Modulation (FiLM) conditions the network continuously on the resolving-power ratio, enabling one model to adapt across slit widths. We further introduce Adaptive RELA, which replaces fixed-kernel rank enhancement with ratio-conditioned multi-scale depthwise convolutions and uses a learnable attention temperature to adjust selectivity with degradation severity. Training data spanning multiple slit configurations are generated using a physics-grounded separable point-spread-function model verified against measured optical data at 48.3 dB accuracy. The resulting model achieves 34-40 dB PSNR across configurations, whereas unconditioned multi-slit training collapses to 24.3 dB and per-slit specialists lose 4-9 dB outside their training setting. It also generalizes smoothly to unseen intermediate configurations without interpolation artifacts. Ablations show that FiLM resolves configuration ambiguity, global linear attention captures long-range directional correspondences, and adaptive temperature yields an additional 2 dB in the challenging near-isotropic regime, where complementary signals are weak.