深度学习驱动的安德森局域化光纤细胞成像技术

115 浏览量更新于2024-08-27 收藏 4.21MB PDF 举报

本文主要探讨了一种基于深度学习的光纤成像系统，该系统利用安德森局域化光学纤维实现了实时无artifact的细胞图像传输。这种创新技术将传统的光纤通信与人工智能相结合，能够在长达一米的光纤中保持细胞图像的高质量，即使在传输过程中遇到温度变化（例如几度的弯曲或加热到50°C的光纤段）也能保持高保真度。系统的核心组成部分是使用了一个深卷积神经网络（Deep Convolutional Neural Network, DCNN），它负责图像的重建过程。训练数据来源于一个使用直光纤在室温（约20°C）下的实验设置，尽管该网络是在特定条件下训练的，但它展现出出色的泛化能力，可以直接应用于通过光纤传输的细胞图像恢复。这种灵活性使得该系统不仅适用于实验室环境，也具有实际应用价值，如远程显微成像或在生物医学设备中的远程监控。文章重点介绍的关键技术包括： 1. **安德森局域化光纤**：这是一种特殊的光纤设计，其结构能够使光信号在传输过程中表现出类似 Anderson 障碍的散射行为，从而限制了光的传播，避免了传统光纤中常见的散射和损耗问题，这对于实现远距离、高清晰度的细胞成像至关重要。 2. **非相干LED光源**：作为照明手段，非相干光源提供了稳定的光源，减少了图像中的噪声和失真，有利于深度学习算法的准确处理。 3. **深度学习算法**：DCNN 的运用使得系统能够在复杂的光纤传输条件下自动学习并适应不同的图像特征，通过训练后的模型，可以实时恢复出经过光纤传输后的细胞图像，提高了图像质量和成像速度。 4. **温度适应性**：实验表明，经过适当训练的网络能够处理不同温度下的光纤传输，这显示了系统的鲁棒性和实用性，可以在各种环境条件下稳定工作。 5. **远程成像**：系统能够捕获几毫米远离光纤前端的细胞图像，拓宽了成像范围，对于观察远离样本区域的细胞动态或者在微流控设备中的应用具有重要意义。这篇研究为光纤成像技术开辟了新的可能，展示了深度学习在解决复杂光学传输挑战中的潜力，为生物医学、远程监测等领域提供了强大的工具。

this system can reach up to several millimeters without any dis-

tal optics. In addition, we show that the image reconstruction

process is remarkably robust with regard to external perturba-

tions, such as temperature variation and fiber bending. Last

but not least, the transfer-learning capability of the new system

is confirmed using cells of different morphologies and classes

for testing. The work presented here introduces a new platform

for various practical applications, such as biomedical research

and clinical diagnosis. The system performance of the Cell-

DCNN-GALOF is superior to state-of-the-art systems. It is also

a new cornerstone for imaging research based on waveguide de-

vices using transverse Anderson localization.

2 Methods

The experimental setup and details of DCNN are shown in

Fig. 1. The GALOF used here is fabricated using the stack-and-

draw method. Silica capillaries with different diameters and air-

filling fractions are fabricated first. The outer diameter of the

silica capillaries ranges from about 100 to 180 μm, and the ratio

of inner diameter to outer diameter ranges from 0.5 to 0.8. To

make a preform, capillaries are randomly fed into a silica jacket

tube. In the following steps, the preform is drawn to canes with

an outer diameter around 3 mm. Finally, the cane is drawn to the

GALOF with the desired size. The SEM image of the GALOF

cross-section is shown in Fig. 1(a).

In Fig. 1(a), the light source is an LED with a center wave-

length of 460 nm. An 80-cm long GALOF sample is utilized.

The diameter of the disordered structure is about 278 μm,and

the air–hole-filling fraction in the disordered structure is

∼28.5%.

The numerical aperture (NA) of the GALOF, based

on far-field emission angles, is measured to be ∼0.4; see Fig. S5

in the Supplementary Material. The temperat ure of a GALOF

segment can be raised by the heater underneath. A 10-mm-long

section in the middle of the GALOF is heated. We use fixed

stained cell samples in all of our experiments. The images of

cell samples are magnified by a 10× objective (NA = 0.3) and

split into two copies sent into a reference path and a measure-

ment path, respectively. The cell samples are scanned both ver-

tically and horizontally with steps of 5 μm to obtain training,

validation, and test data sets. In the reference beam path,

the image is further magnified by a 20× objective (NA ¼ 0.75)

and recorded by CCD 1 (Manta G-145B, 30 fps) after passing

through a tube lens. In the measurement path, the image is trans-

ported through the 80-cm -long GALOF and then projected onto

CCD 2 (Manta G-145B, 30 fps) by the same combination of

a 20× objective and tube lens. The reference images are labeled

as the ground truth. Both reference and raw images are 8-bit

grayscale images and are cropped to a size of 418 × 418 pixel s.

Figure 1(b) shows that experiments are performed for both

straight GALOF and bent GALOF. To bend the fiber, the input

end of the GALOF is fixed, whereas the output end of the

GALOF is moved by an offset distance. The amount of bending

is quantified by the offset distance from the end of the bent fiber

to the position of the straight fiber (equal to the length of the

dashed line). The relation between the offset distance d and

the corresponding bending angle of the fiber θ is given by d ¼

L½1 − cosðθÞ∕θ where L is the total length of the GALOF.

Figure 1(c) shows the detailed structure of the DCNN. The

raw image, which is resized to 420 × 420 using zero paddin gs,

is the input layer. The input layer is decimated by five down-

sampling blocks (blue and black arrows) to extract the feature

maps. Then five up-sampling blocks (white arrows) and one

convolutional block (yellow arrow) are applied to reconstruct

the images of cell samples with a size of 418 × 418. To visualize

the image reconstruction process, some sample feature maps are

shown in Fig. S6 in the Supplementary Material. The skip con-

nections (dark green arrows) pass feature information from fea-

ture-extraction layers to reconstruction layers by concatenation

operations. The mean absolute error (MAE)-based loss metrics

are calculated by comparing the reconstructed images with the

reference images. The MAE is defined as jI

rec

− I

ref

j∕ðwhÞ,

where I

rec

, I

ref

, w, and h are the reconstructed image intensity,

the reference image intensity, the width, and the height of the

images, respectively. The parameters of the DCNN are opti-

mized by minimizing the loss. Detailed block operation dia-

grams corresponding to the respective arrows are shown on

the right side of Fig. 1(d) (BN, batch normalization; ReLU, rec-

tified linear unit ; Conv, convolution; D-Conv, dilated convolu-

tion; T-Conv, transposed convolution; concat, concatenation).

The Keras framework is applied to develop the program code

for the DCNN. The regularization applied in the DCNN is de-

fined by the L2-norm. The parameters of the DCNN are initial-

ized by a truncated normal distribution. For both training and

evaluation, the MAE is utilized as the metric. The Adam opti-

mizer is adopted to minimize the loss function. During the train-

ing process, the batch size is set at 64 and the training is run

through 80 epochs with shuffling at each epoch for all of the

data shown in this paper. The learning rate is set at 0.005. Both

training and test processes are run in parallel on two GPUs

(GeForce GTX 1080 Ti).

3 Results

3.1 Imaging of Multiple Cell Types

To demonstrate the imaging reconstruction capability, two dif-

ferent types of cells, human red blood cells and cancerous

human stomach cells, serve as objects. By scanning across dif-

ferent areas of the cell sample, we collect 15,000 reference and

raw images as the training set, 1000 image pairs as the validation

set, and another 1000 image pairs as the test set for each type of

cell. During the first data acquisition process, the GALOF is

kept straight and at room temperature of about 20°C. The im-

aging depth is 0 mm, meaning that the image plane is located

directly at the fiber input facet. The training data are loaded into

the DCNN [see Fig. 1(c) for DCNN structure] to optimize the

parameters of the neural network and generate a computational

architecture that can accurately map the fiber-transported im-

ages to the corresponding original object. After the training pro-

cess, the test data are applied to the trained model to perfor m

imaging reconstruction and evaluate its performance using

the normalized MA E as the metric. In the first round of experi-

ments, we train and test each type of cell separately. With a

training data set of 15,000 image pairs, it takes about 6.4 h

to train the DCCN over 80 epochs on two GPUs using a per-

sonal computer. The accuracy improvement curves for both

training and validation processes over all 80 epochs are pro-

vided in Fig. S1 in the Supplementary Material. After training,

the reconstruction time of a single test image is about 0.05 s.

Figure 2 shows some samples from the test data set. In

Figs. 2(a)–2(c), reference images, raw images, and recovered

images of three in succession collected and reconstructed im-

ages of human red cells are shown, whereas in Figs. 2(d)–2(f),

three images of cancerous stomach cells are presented.

Comparing the reference images with the reconstructed images,

Zhao et al.: Deep-learning cell imaging through Anderson localizing optical fiber

Advanced Photonics 066001-3 Nov∕Dec 2019

•

Vol. 1(6)

剩余11页未读，继续阅读

weixin_38704011

粉丝: 3
资源: 947

深度学习驱动的安德森局域化光纤细胞成像技术

医疗影像数据集压缩包dl-medical-imaging介绍

利用coherent-diffraction-imaging技术的游戏开发

Modelhub: 深度学习模型库与统一API的创新应用

Presentation-2021-04-21-deep-learning-medical-imaging:深度成像2021学校中的不确定性和贝叶斯神经网络演示

Large-Scale-Multi-Class-Image-Based-Cell-Classification-with-Deep-Learning:具有深度学习的大规模多类基于图像的细胞分类

Learning-based lensless imaging through optically thick scattering media

Self-Supervised-Learning-in-Medical-Imaging

Deep-Learning-with-PyTorch.rar

颜色分类leetcode-Awesome-Machine-Learning-in-Biomedical-Healthcare-Imaging:

matlab人脸匹配代码-deep-imaging:深度成像

最新资源