研究目的
To develop a deep learning-based method for optical sectioning of wide-field images that requires minimal training data and offers lower noise levels and higher imaging depth compared to traditional methods.
研究成果
The deep learning-based method for optical sectioning achieves similar background suppression and resolution to traditional methods but with lower noise and higher imaging depth. It offers a cost-effective and convenient solution for high-throughput optical sectioning, with potential for real-time imaging applications.
研究不足
The method requires a well-aligned pair of wide-field and optically-sectioned images for training. The risk of overfitting increases with the size of the sub-images used for prediction, necessitating more training data.
1：Experimental Design and Method Selection:
A convolutional neural network (CNN) was designed to approximate the inverse image formation process for optical sectioning. The network includes multiple convolutional layers with parametric rectified linear units (PReLU) as activation functions.
2：Sample Selection and Data Sources:
Training data consisted of one pair of wide-field and corresponding optically-sectioned images acquired using a homemade structured illumination microscope (SIM).
3：List of Experimental Equipment and Materials:
A homemade SIM with a 20× objective (XLUPLFLN-W, NA 1.0, Olympus Inc., Japan), FluoSpheres Carboxylate-Modified Microspheres (Molecular Probes), and a TITAN Xp (NVIDIA) GPU for training.
4：0, Olympus Inc., Japan), FluoSpheres Carboxylate-Modified Microspheres (Molecular Probes), and a TITAN Xp (NVIDIA) GPU for training.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: The CNN was trained using a pair of images, then used to predict optically-sectioned images from new wide-field images. The process involved splitting images into sub-images, training the CNN, and stitching predicted sub-images together.
5：Data Analysis Methods:
Performance was evaluated by comparing predicted images to those obtained with SIM and RL deconvolution, using peak signal to noise ratio (PSNR) and signal-to-background ratio (SBR).
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
