Deep-Learning Generation of High-Resolution Images of Live Cells in Culture Using Tri-Frequency Acoustic Images

Ultrasound microscopy is the only technique that has the ability to monitor live-cell morphology over a long period of time without causing any damage to the cells, but its longer wavelength prevents one from obtaining high-resolution cell images. Here, we propose a deep-learning (DL) method for generating high-resolution acoustic images. By preparing datasets consisting of many pairs of acoustic and optical-microscope images for the same cells and training them, a high-resolution image comparable to optical microscopy is generated from an acoustic image. Importantly, the most accurate images are generated when three-layer (RGB) images containing not only high-frequency (approximately 180 MHz) images but also lower-frequency (approximately 100 MHz) images are used as the input images, which is attributed to enhanced acoustic absorption in the nucleus because the nucleus resonates in this low-frequency band. The DL scheme with the tri-frequency image input is applied to human mesenchymal stem cells and human induced pluripotent stem cells, and the high image-generation capability is demonstrated. As a result, high-resolution acoustic microscopy images are obtained for the same cells for over 24 h, without the typical cell damage encountered using optical imaging.