INFORMATION TECHNOLOGY

Unsupervised learning breaks through fiber imaging barriers


Guide

A collaborative team from MIT and the University of Central Florida recently developed an unsupervised learning fiber imaging system based on a cyclically generative adversarial network architecture and Anderson local microstructured fibers.

The system realizes high-quality and highly robust color biological image transmission based on label-free small data model training. The image transfer process also showed high universality for different kinds of biological samples that were not included in the training data. This new fiber optic imaging system will be used in medical diagnosis and basic biological research.

The article was published in Light: Science & Applications under the title “Unsupervised Full-color Cellular Image Reconstruction through Disordered Optical Fiber”, with Xiaowen Hu as the first author and Jian Zhao as the corresponding author of this article.

In clinical applications and basic biology research, doctors or researchers often face special imaging environments, such as the cavity of human organs or the cerebral cortex of experimental mice.

Traditional microscopy imaging systems are difficult to directly obtain images of the walls of living organ cavities or the cerebral cortex due to their large size and inability to penetrate deep into living organisms. In contrast, the high flexibility and small spatial size of optical fibers (most fibers are between 125 microns and 1 mm in diameter) allow fiber-based imaging systems to capture and transmit biological images deep into living organisms. This makes it have great application value in clinical diagnosis and basic science. However, solutions based on traditional algorithms and commercial optical fibers face problems such as poor image quality and low robustness of the image transmission process.

In recent years, the resurgence of artificial intelligence has also led to the development of optical fiber imaging system solutions based on deep learning. These solutions initially overcome the shortcomings of traditional systems and show great potential. Despite tremendous progress, existing programmes rely mostly on supervised learning. Supervised learning requires a large amount of strictly labeled image data for model training, which puts forward high requirements for data acquisition, system design and system calibration of optical fiber imaging systems. The corresponding disadvantages are that the time cost of data acquisition is increased, the complexity of the system structure and experimental process is improved, and the calibration process and times of the system are also put forward higher requirements. These problems will hinder the practical application of fiber optic imaging systems based on supervised learning in the future.

Recently, Dr. Jian Zhao from the Picowl Institute for Learning and Memory at the Massachusetts Institute of Technology, Dr. Xiaowen Hu from the School of Optics and Photonics at the University of Central Florida (now working at ASML), Professor Axel Schülzgen and related team members have developed a fiber imaging solution based on unsupervised learning (Restore-CycleGAN-GALOF), breaking through the technical bottleneck of fiber imaging under the framework of supervised learning.

Compared with the fiber imaging solution based on supervised learning, the Restore-CycleGAN-GOF scheme has made innovations in two directions: deep learning algorithm and fiber imaging system.

In terms of algorithmsBased on the unsupervised learning cyclic generative adversarial network (CyleGAN) architecture, a Restore-CycleGAN model adapted to fiber imaging was developed and used for image reconstruction.

In terms of fiber optic imaging systemsThe new scheme is based on Anderson Local Area Fiber (GALOF) with a random structure to develop two imaging systems, transmission and reflection.

Through the organic combination of the Restore-CycleGAN algorithm and the GALOF fiber imaging system, the new Restore-CycleGAN-GALOF achieves the following breakthroughs:

1. Based on the single unlabeled small dataset model training, high-quality color biological image optical fiber transmission is realized;

2. Realize the image transmission process with high robustness;

3. Realize a highly ubiquitous deep learning-based imaging process.

The Restore-CycleGAN-GALOF scheme only needs to use 1000 sets of label-free image data for model training, which can achieve high-quality color biological image transmission. Compared with previous supervised learning-based fiber imaging schemes, the new method reduces the training data by at least ten times. More importantly, for the new scheme, a single model training can achieve high-quality image transmission under different experimental conditions within a certain range. The successful implementation of this single-time, label-free, small-data set model training can significantly improve the imaging speed, simplify the experimental system and experimental steps, and better meet the needs of practical applications.

Based on the unsupervised learning described above, the Restore-CycleGAN-GALOF imaging system demonstrates high-quality color image transmission to different biological samples. This high-quality color image transmission process shows high stability for large fiber bends (60 degrees called mechanical bending) and varying sample distances (0 mm – 6 mm). The high stability of this image transmission is ideal for imaging complex living organs. In addition to superior robustness, the Restore-CycleGAN model has been shown to be able to reconstruct images of different kinds of biological samples that have never appeared in the training dataset with high quality, showing strong universality. It is worth noting that this universality is achieved under small sample training data, while previous imaging systems based on supervised learning often need to achieve limited generalization through large-sample training. Restore-CycleGAN proves that unsupervised learning can improve the universality of fiber image reconstruction without significantly increasing the training dataset. This universality helps further compress the time for data acquisition and simplify the required experimental setup and steps in biomedical imaging applications.

Figure 1: Restore-CyleGAN-GALOF schematic

Figure 2: Example of Restore-CyleGAN-GALOF bioimaging results

The future direction will be to move further from proof-of-concept to more practical endoscopic systems. This will require further improvement of GALOF’s structural parameters and optimization of existing optical system designs for more compact miniaturized imaging systems, while further developing the corresponding unsupervised learning algorithms. On this basis, relevant animal imaging experiments will help to advance its practical application. In the future, imaging systems based on the Restore-CylceGAN-GOF concept may be applied to different application scenarios. Especially in clinical diagnosis and basic brain science research, unsupervised learning fiber imaging may play a huge role. (Source: China Optics WeChat public account)

Related paper information:https://doi.org/10.1038/s41377-023-01183-6

Special statement: This article is reproduced only for the need to disseminate information, and does not mean to represent the views of this website or confirm the authenticity of its content; If other media, websites or individuals reprint and use from this website, they must retain the “source” indicated on this website and bear their own legal responsibilities such as copyright; If the author does not wish to be reprinted or contact the reprint fee, please contact us.



Source link

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button