Facilitating real-time LED-based photoacoustic imaging with DenP2P: An optimized conditional generative adversarial deep learning solution

Photoacoustic imaging (PAI) benefits from the optical absorption contrast of the tissue while achieving greater depth information with ultrasound resolution than the other optical imaging platforms. The recent advancement of moderate pulse width LED (e.g., Acoustic-X) illuminating devices makes PAI more affordable, mobile, and fast with a trade-off for low illumination energy, leading to low signal-to-noise-ratio (SNR) images, which are averaged over time to get high SNR images. Signal quality can be improved by traditional noise removal algorithms, but deep learning models outperform non-learning methods. Although the most widely used U-Net architecture removes noise, it compromises the structural similarity, introduces blur, and causes edge artifacts. To mitigate those issues, we explored a gamut of architectural options of the Pix2Pix-based conditional generative adversarial network (cGAN) analyzing objective functions, optimizers, activation, and normalization layers. The optimized denoising cGAN model based on Pix2Pix architecture (DenP2P) is tested with a variety of out-of-class biological test data such as in vivo mouse tumor, ex vivo kidney, heart, and liver spatially different from training examples. The network is also tested for noise distribution type invariancy concerning Gaussian white, Poisson, speckle, and salt and pepper noise. The frequency domain’s magnitude spectrum analysis explains less blurring for the generated outcomes compared than U-Net. In addition, the persistence homology diagrams (birth, death, and lifetime) underscored DenP2P’s efficacy in diminishing noisy topological features, fostering the emergence of stable and resilient structures. Overall, the optimized DenP2P model generates high SNR images with appreciable peak SNR and structural similarity index.