An Optical Fiber Imaging System Based on Unsupervised Learning Presented by Researchers

Fiber-optic imaging methods enable in vivo imaging within hollow organs or tissues otherwise inaccessible to space-free optical techniques, and play a vital role in clinical practice and basic research, such as endoscopic diagnostics and deep brain imaging.

Recently, fiber-optic imaging methods based on supervised learning have gained popularity due to their superior performance in recovering high-resolution images from degraded fiber-delivered images or even fuzzy speckle patterns. Despite their success, these approaches are mainly limited by their requirements for accurate paired labeling and large training datasets.

Difficult training data requirements lead to time-consuming data acquisition, complex experimental design, and laborious system calibration processes, which makes it difficult to meet practical application needs.

In a recent post on Light: science and applicationsDr. Jian Zhao of MIT’s Bequewer Institute for Learning and Memory, Dr. Xiaowen Hu and Dr. Axel Schulzgen of the College of Optics and Photonics (CREOL) at the University of Central Florida, and their colleagues demonstrated an unsupervised learning-based fiber-optic imaging system.

This system integrates a dedicated generative cycle adversarial network (CyleGAN), called Restore-CycleGAN, with Glass-Air Anderson’s Local Optical Fiber (GALOF). The Restore-CycleGAN implementation removes the limitations of labeled training data, yet maintains high-quality imaging recovery, while the unique physical properties of the GALOF modes support extremely robust, high-resolution imaging operations and ensure successful execution of unpaired imaging training.

Due to cross-promotion between the learning algorithm and the optical hardware, the Restore-CycleGAN-GALOF method achieves robust and virtually artifact-free transmission of full-color biological images through a meter-long optical fiber using a simple one-shot training process using a small training dataset of 1000 pairs of Images only, with no associated training imaging data required. Training data volume is reduced by about ten times compared to previously reported supervised learning methods.

The Restore-CycleGAN-GALOF method demonstrated the ability to transmit high-resolution, color images of several biological samples, including human and frog blood cells, human eosinophils, and human gastric cancer cells, under both transmission and reflectance imaging modes.

Moreover, this imaging process showed resilience against strong mechanical fiber bending of 60° and large variations in working distance of up to 6 mm. Remarkably, the Restore-CycleGAN-GALOF method produced high-accuracy predictions for test data that were never included in the training process, indicating strong generalizability in the small data system.

Despite the superior performance of Restore-CycleGAN-GALOF, the system design and experimental operation are relatively simple. The scientists summarized the importance of their imaging method: “Access to the far end of fiber devices and collection of sufficient training data are challenging in practical applications. The unique environments of biological organs or tissues pose additional difficulties for robust image transfer.”

However, our Restore-CycleGAN-GALOF method requires only a small amount of training data and eliminates the need to pair image features. In a small data system, this method ensures highly robust, generalizable, robust color imaging. As a result, it is particularly suitable. Better to meet various practical biomedical applications.

“Our technologies are expected to lay the foundation for the next generation of fiber-optic imaging system. Our future research will focus on developing practical endoscopy systems and conducting tests for relevant biomedical applications. We aspire to advance medical diagnostics and basic biological research through our methodology,” the scientists added.