High-resolution imaging in acoustic microscopy using deep learning

Acoustic microscopy is a cutting-edge label-free imaging technique to visualize surface and internal structure of industrial and biological specimens. Focusing high-frequency acoustic waves to the object followed by detection of echo signals results in the final acoustic image. However, resolution of the acoustic image depends on the signal to noise ration, step size, and frequency of the transducer. In this work, we propose a deep learning enabled high-resolution imaging in acoustic microscopy. The approach is based on training various generative adversarial networks (GAN) to transform low-resolution transducer limited images into super-resolved images. Total five different type of GAN model i.e., SRGAN, ESRGAN, IMDN, DBPN-RES-MR64-3, and SwinIR is used to show 2.5 times resolution enhancement in acoustic images. The performance of the trained model is evaluated by calculating PSNR and SSIM between the network predicted and ground truth images. SwinIR provided an average SSIM AND PSNR of 0.92 and 35, respectively. A biological sample was tested on the model and it resulted in a SSIM score of 0.8778 and PSNR score of 32.93. Our acoustic microscopy +GAN framework is applicable for various industrial applications such as electronic manufacturing, micro-structural characterization of material and other biomedical applications in general.