Abstract
Spectroscopic photoacoustic (sPA) imaging provides critical functional information for medical diagnostics by identifying tissues based on their unique optical absorption spectra. However, its clinical utility is often limited by high noise levels and the lack of large-scale "ground truth" datasets required for traditional deep learning denoising. This paper presents SPADE (Spectroscopic Photoacoustic Denoising), a hybrid framework that achieves over 15 dB SNR improvement using a "zero-shot" approach. By combining analytical filtering with data-free deep learning, SPADE preserves spectral integrity without the need for pre-trained models, making it highly adaptable to diverse clinical environments.
Introduction
Spectroscopic photoacoustic (sPA) imaging is a non-invasive technique that uses laser pulses to generate ultrasound waves, allowing for the visualization of chromophores like hemoglobin and gold nanoparticles. Despite its potential, sPA suffers from low signal-to-noise ratios (SNR) due to electronic noise and acoustic interference. Traditional denoising often requires frame averaging, which compromises temporal resolution, or deep learning models that require massive amounts of labeled training data—a resource rarely available in clinical settings. SPADE addresses these challenges by learning to denoise directly from the noisy data itself.
The SPADE Framework
The SPADE framework operates in three distinct stages to maximize noise reduction while maintaining the spectral correlations necessary for accurate tissue identification.
1. Spectral Domain Data Re-assembly (SDDR)
To ensure the denoising algorithm accounts for spectral relationships, SPADE first reassembles the 3D data cube (spatial dimensions $H \times W$ and wavelength dimension $\lambda$) into a 2D pseudo-image. This process, termed SSDR, groups pixels from the same spatial location across different wavelengths. This re-assembly allows the subsequent filters to "see" and preserve the spectral signatures essential for unmixing.
2. Analytical Denoising (BM3D)
The framework utilizes a modified Block-Matching and 3D filtering (BM3D) algorithm as an initial stabilizer. BM3D is a classical analytical method that excels at reducing Gaussian noise. In SPADE, it provides a cleaner baseline for the learning-based stage, preventing the neural network from over-fitting to extreme noise artifacts.
3. Zero-Shot Noise2Noise (ZS-N2N)
The final stage is a "zero-shot" learning component. Unlike standard Noise2Noise models that require pairs of noisy images, ZS-N2N uses a single noisy frame. By applying specific convolution masks to downsample the image into pairs with independent noise but shared underlying signals, the network learns to predict and subtract noise in real-time for that specific dataset.
Experimental Validation and Results
The authors validated SPADE across four scenarios: numerical simulations, physical wire phantoms, ex vivo swine cardiac tissue, and in vivo mouse tumor models.
- Performance: In simulations, SPADE achieved an SNR improvement of over 15 dB. In ex vivo studies, it reached an SNR of 43.05 dB, nearly matching the 43.22 dB achieved by averaging 64 frames, but requiring only a single frame.
- Spectral Fidelity: A key metric in sPA is the spectral correlation coefficient ($R$). SPADE maintained $R > 0.8$ even under high-noise conditions, ensuring that the "color" of the tissue—and thus its functional identity—remained accurate.
- Clinical Utility: In vivo tests on mice with prostate cancer showed that SPADE effectively eliminated false-positive signals from contrast agents, providing a clearer map of tumor boundaries compared to standalone analytical or learning-based methods.
Conclusion
SPADE represents a significant step toward practical, high-fidelity sPA imaging. By eliminating the dependency on pre-trained datasets, it offers a hardware-agnostic solution that can be deployed across different clinical settings. While current processing times (approx. 70 seconds per multi-wavelength image) are a hurdle for real-time applications, future optimizations for GPU acceleration and non-Gaussian noise handling promise to bring this "data-free" enhancement framework into the operating room.
Summary updated