Amongst the several biomedical imaging modalities, Photoacoustic imaging stands out due to its advantage of providing optical contrast at ultrasound resolution from deeper tissues. The optical illumination is traditionally provided by the nanosecond-pulse width lasers, but they are costly, bulky, and non-portable. Light Emitting Diode-based systems can circumvent all these issues, but they deliver low-energy that brings forth another problem of low signal-to-noise-ratio (SNR) images. Averaging several frames at the same cross-section over time removes the noise, but real-time dynamic functionalities might not be captured. The tradeoff between SNR and real-time acquisition can be mitigated with a downstream noise removal algorithm. The traditional algorithms are not efficient and require prior knowledge about the noise type distribution for which deep learning-based architectures such as U-Net and generative adversarial network (GAN) are implemented. One of the issues of these supervised networks is the requirement of paired training input-label dataset which is highly cumbersome to capture or sometimes is unavailable. The pixel-wise correspondence will act as a pre-processing overburden for acquiring training data. To mitigate this issue, we implemented a Cycle-consistent GAN denoising (DenCyc-GAN) algorithm which works on unpaired training data. We compared our network’s outputs with other traditional non-learning and deep learning network and found that our network performed similar to the supervised networks with respect to image quality metrics such as Peak SNR and structural similarity index.
|