1. How can deep learning (DL) be applied in quantitative phase microscopy (QPM) for phase image prediction?
Deep learning (DL) can be applied in quantitative phase microscopy (QPM) for phase image prediction by utilizing a U-net network trained with phase-intensity image pairs. In this approach, the phase images are obtained using an off-axis DHM configuration and a phase-type spatial light modulator (SLM) to generate a series of phase samples. The trained network can accurately predict the phase information from a single defocused bright-field intensity image. This DL-based approach eliminates the need for recording multiple defocused intensity images, reducing experimental time and costs. Additionally, DL has been introduced into DHM to address phase recovery and aberration compensation problems, enhancing the performance of existing imaging approaches. Overall, DL can extend the capabilities of QPM by enabling the conversion of images between different modalities and improving the accuracy of phase image prediction.
read more
2. What is the role of the U-net in DLQPM imaging?
The U-net in DLQPM imaging serves as a nonlinear operator that links the relation between defocus intensity and phase information. It is expressed with a U-net network, which consists of an encoder and a decoder. The encoder performs feature extraction through five repetitions, each containing two 3x3 convolution layers, batch normalization, ReLU function, and a residual block. The decoder performs up-sampling through transpose convolution and two 3x3 convolutions, batch normalization, ReLU, and a residual block. The U-net network is trained with throunds of intensity-phase data pairs, allowing it to predict phase images from defocus intensity images.
read more
3. How is U-net trained with I d -ph data pairs?
U-net is trained with I d -ph data pairs by capturing defocus brightfield intensity images I d and off-axis holograms I h using a CCD. The in-focus phase images ph are calculated using a DHM recovery algorithm and digitally propagated to the image plane. The acquired I d and ph are cropped to a size of 256 x 256 pixels, with the phase images serving as the ground truth input for the neural network. The training and testing of the network model are performed using the Pytorch framework on a PC with an Intel Core processor and NVIDIA GeForce GTX2060 GPU.
read more
4. How is phase accuracy verified in the proposed neural network?
Phase accuracy in the proposed neural network is verified by loading handwritten numerals to the SLM and testing the DLQPM network. A total of 3,656 images of numerals were used, with 3,500 for training and 156 for testing. Off-axis holograms and defocused intensity images were recorded for each sample. The reconstructed phase image from the neural network is compared to the ground-truth phase image reconstructed by off-axis DHM using the Structural Similarity Index Measure (SSIM). The SSIM value between the neural network output image and the Ground Truth is 0.965, indicating high accuracy. Random phase patterns generated by SLM were also tested, showing the neural network's ability to accurately reconstruct phase images with a standard deviation of 0.15 rad for samples with a peak-to-valley (P-V) value around 1.5 rad. The error may be caused by speckle noise and phase fluctuation during DHM imaging. Overall, the neural network can accurately recover phase information from bright-field defocus intensity images.
read more