Deep learning-enhanced LED-based photoacoustic imaging

Abstract

Photoacoustic imaging holds promise in wide range of clinical and preclinical applications. Since photoacoustic imaging can be implemented in a conventional ultrasound scanner by adding light illumination, it is straight forward to realize dual-mode imaging offering complementary contrast. We recently developed an LED-based photoacoustic and ultrasound imaging system (AcousticX) with unprecedented 2D and 3D functional and structural imaging capabilities. Pulse energy offered by our LED arrays is orders of magnitude lower than conventional lasers and we perform frame averaging to keep up with the SNR, reducing the display frame rate. Even though the pulse repetition frequency of our LED arrays is 4 KHz, image frame rate we can achieve is limited by the large number of frame averages used to improve SNR. In this work, we present a deep learning-based approach to reduce the frame averaging in LED-based photoacoustic imaging without compromising the SNR. We have used convolutional neural network (U-Net) model in deep learning for improving the images with less averaging. When compared with traditional denoising methods, deep learning enables us to optimize parameters through network training. We used images from various other photoacoustic imaging systems with higher laser energy and broadband ultrasound transducers, which can generate PA images with high resolution and SNR with minimal or no averaging as training data. We validate our algorithm using LED-based photoacoustic images of phantoms utilizing Indocyanine green and methylene blue as contrast agents. In all cases, we achieved improvement in the SNR by denoising the images with lesser averaging, thereby increasing the framerate. Results demonstrate the potential of deep learning algorithms in improving temporal resolution and SNR in LED-based photoacoustic imaging.