Introduction
Photoacoustic Microscopy (PAM) is a non-invasive imaging modality that provides high-contrast visualization of microvasculature by combining optical excitation with ultrasonic detection. It is particularly effective for imaging blood vessels in the brain. However, PAM systems face a fundamental trade-off between imaging speed and spatial resolution. High-resolution (HR) images require dense mechanical scanning, which is time-consuming. Conversely, accelerating the process through sparse sampling results in low-resolution (LR) images with significant artifacts and loss of structural detail.
While deep learning-based super-resolution (SR) has shown promise in general computer vision, applying it to PAM is challenging due to hardware-specific distortions. This research introduces a novel SR framework that incorporates "scanning priors"—knowledge of the hardware's mechanical behavior—to reconstruct high-fidelity vascular images from sparsely sampled data.
The Challenges of PAM Scanning
Standard SR models often fail on PAM data because they do not account for the physical realities of the scanning process. Two primary issues complicate the reconstruction:
- Scanner Wobbling: Mechanical vibrations during scanning cause misalignments between adjacent rows of pixels, leading to a "jagged" appearance in vessels.
- Scanning Geometry: Rotational scanning, often used to cover large areas quickly, introduces unique artifacts. When converting data from polar to Cartesian coordinates, "zero-point masks" or areas of missing data can appear, which general AI models struggle to interpret correctly.
Proposed Methodology
The authors propose a comprehensive pipeline that integrates hardware-aware preprocessing with a state-of-the-art Transformer architecture. The framework is designed to ensure that reconstructed images are not just visually sharp but also anatomically accurate for downstream medical tasks.
1. Registration Module
To address scanner wobbling, the framework includes a preprocessing registration step. Using the Speeded Up Robust Feature (SURF) algorithm, the system separates odd and even scanning rows, aligns them to correct for mechanical shifts, and then recombines them. This ensures structural continuity before the image enters the neural network.
2. Swin-Transformer Backbone
The core of the SR model is based on SwinIR, a Transformer-based architecture. Unlike traditional Convolutional Neural Networks (CNNs) that focus on local pixel neighborhoods, Transformers utilize self-attention to capture long-range dependencies. This is crucial for PAM, as it allows the model to understand the global connectivity of blood vessels across the entire field of view.
3. Gradient-Based Patch Selection
Training efficiency is improved by prioritizing "vessel-rich" areas. Instead of random sampling, the system calculates gradient scores for image sub-windows. Areas with high gradients—indicating sharp vessel edges—are more likely to be selected for training, forcing the model to focus on the most difficult and important features.
4. Scanning Consistency Loss ($L_{sc}$)
A novel loss function, the Scanning Consistency Loss, was developed to handle rotational scanning artifacts. It enforces gradient consistency between the reconstructed image and the ground truth specifically within the transformed image space, helping the model ignore the inherent "zero-point" artifacts of the rotational geometry.
Experimental Results
The framework was evaluated using synthetic mouse brain microvasculature data and real-world clinical datasets. It was compared against several state-of-the-art methods, including EDSR, DAT, and FD-Unet.
- Quantitative Performance: On synthetic data at a $2\times$ scale factor, the model achieved a PSNR of 42.935 dB and a DICE score of 97.178%. The high DICE score is particularly significant, as it indicates that the reconstructed vessels are highly accurate for automated segmentation.
- Real-World Generalization: When tested on noisy, real-world raster-scanned images, the model maintained a DICE score of 71.461%, outperforming both Bicubic interpolation and other deep learning models. This demonstrates the framework's robustness across different hardware configurations.
- Visual Fidelity: Qualitative analysis showed that the proposed method significantly reduced broken vessel connections and artifacts compared to general-purpose SR models, producing much cleaner error maps in vascular regions.
Conclusion
By incorporating hardware-specific scanning priors into a Transformer-based architecture, this research provides a robust solution for high-speed, high-resolution PAM imaging. The integration of a mechanical registration module and a specialized scanning consistency loss allows the model to overcome the physical limitations of the scanning hardware. Future work will explore multi-frame super-resolution to leverage temporal information and further enhance image quality in dynamic clinical environments.
Summary updated