|
Multiphoton microscopy is a very important and powerful technique in the application of in-vivo drosophila brain imaging. However, the photon budget, which is compromising between imaging speed, spatial resolution and accumulating time, remains a huge challenge in biological compatibility. Rapid dual-resonant volumetric multiphoton microscopy combines a tunable acoustic gradient (TAG) lens and a resonant mirror, which can achieve up to 8 kHz frame rate and hundreds of hertz volumes per second temporal resolution. Adaptively sampling each laser pulse by an embedded field programmable gate array (FPGA) with up to 80 MHz pixel rate enables efficient but restricted signal accumulation. This study has developed a generative neural network to restore images from degradation of low signal-to-noise ratio (SNR), missing pixels and pattern residual. A series of training data by adding noise and Lissajous scanning path into the known fluorescent bead images were used to pretrain the model for the rapid dual resonant volumetric multiphoton microscopy system. Experimental results verify that the axial distortion and the image resolution of noisy fluorescent bead images can be effectively restored with the deep-restoration neural network. The mushroom body (MB) of drosophila brain which contains thousands of Kenyon cells in a 200 × 200 × 100 µm3 volume is utilized to demonstrate the strategy. The deep-restoration rapid dual resonant volumetric multiphoton microscopy image not only maintains 256 × 256 × 128 voxels and ~30 volumes per second, but also significantly improves image quality which is compatible to the ground truth. However, in-vivo imaging is time-varying. Base on in-vitro image datasets, in-vivo images via transfer learning is utilized to enables fast and improved image quality efficiently. Despite of one pulse per pixel, in-vivo drosophila brain imaging with deep-restoration not only keeps the advantage of temporal resolution, but also obtains well image quality.
|
