Neural Network Achieves High-Speed Dynamic Optical Coherence Tomography with 4 OCT Volumes

Dynamic optical coherence tomography (DOCT) promises real-time, high-resolution imaging of biological processes, but conventional full-field swept-source coherence microscopy (FF-SS-OCM) struggles with the massive data volumes generated during dynamic scans. Suzuyo Komeda, Nobuhisa Tateno, and Yusong Liu, from the Computational Optics Group at the University of Tsukuba, alongside Rion Morishita, Xibo Wang, and Ibrahim Abd El-Sadek, have developed a neural-network-based method to overcome this limitation. Their research demonstrates a significant reduction in data size and processing time, achieving high-definition DOCT images from just four OCT volumes, compared to the thirty-two traditionally required. This innovation enables substantially faster imaging, reducing data transfer from 42 GB to 5.3 GB and processing time from four hours to just thirty minutes, representing a major step towards practical, high-throughput dynamic imaging applications.

FF-SS-OCM provides high-definition OCT images, but particularly in DOCT imaging, it results in a significant enlargement of the data size and subsequently long data streaming and processing time, which prevents high-throughput imaging. This work addresses this issue by introducing an NN-based DOCT method that generates high-definition logarithmic intensity variance (LIV)-based DOCT images. The research objective is to accelerate DOCT image processing without compromising image quality, enabling high-throughput imaging applications. This is achieved through the development of a neural network capable of efficiently reconstructing DOCT images from input data, effectively reducing data processing time and computational load.

Summary of the Paper Title: Neural Network-Based High-Speed Volumetric Dynamic Optical Coherence Tomography for Tumor Spheroid Evaluation – I. A. El-Sadek – R. Morishita – N. Tateno – S.

Komeda, S. Makita, and Y. Yasuno present a novel approach to high-speed volumetric dynamic optical coherence tomography (dyc-OCT) that leverages neural networks for the evaluation of tumor spheroids. The work focuses on enabling real-time, label-free imaging and analysis of intratissue dynamics, addressing the need for faster and more accurate assessment of cellular motility and tissue behavior in cancer research. Dynamic OCT is highlighted as a critical tool for studying these processes, with the primary objective being the development of a high-speed volumetric dyc-OCT system enhanced through neural network–based processing.

The proposed system is based on Dynamic Full-Field Swept-Source OCT (DFS-OCT), which combines full-field imaging with swept-source technology to achieve high-speed volumetric acquisition. A U-Net neural network architecture is employed to process raw OCT data and extract dynamic features in real time. The system is evaluated using multicellular tumor spheroids, a standard in vitro model for cancer studies, and dynamic contrast OCT is used to capture subtle temporal changes in tissue properties. A large, labeled OCT dataset is collected to train the network, which is optimized using the Adam optimizer with learning rate decay. Post-processing steps are applied to the network output to improve image contrast and suppress noise.

Experimental results demonstrate that the system achieves high spatial resolution of approximately 1.5 μm and volumetric imaging speeds of up to 20 volumes per second. The neural-network-assisted dyc-OCT approach produces high-quality images with minimal artifacts and enables accurate quantification of intratissue dynamics, including cell motion and necrotic regions within tumor spheroids. Compared with conventional OCT techniques, the proposed method shows significant improvements in both imaging speed and analytical accuracy, offering more detailed insights into cellular behavior and tissue responses to treatment.

In conclusion, the study introduces a promising neural-network-based framework for high-speed volumetric dynamic OCT that substantially enhances tumor spheroid evaluation. By enabling real-time, label-free analysis of intratissue dynamics, the approach has strong potential for advancing biomedical imaging and cancer research. Future work is expected to focus on further optimizing the algorithm and extending its application to a broader range of biological systems.

Neural Networks Accelerate High-Definition DOCT Imaging

Scientists have achieved a significant breakthrough in dynamic coherence tomography (DOCT) imaging by developing a neural-network (NN)-based method for high-speed, high-definition data acquisition. The research, utilizing full-field swept-source coherence microscopy (FF-SS-OCM), addresses the challenge of large data sizes and lengthy processing times inherent in conventional DOCT. Experiments revealed that the team successfully generated high-definition logarithmic intensity variance (LIV)-based DOCT images from just four OCT volumes, a substantial reduction compared to the 32 volumes required by traditional methods. This innovative approach delivers an eight-fold reduction in data size, decreasing measurements from an initial 42 GB to 5.3 GB.

Data transfer times were dramatically improved, falling from 7 minutes to 55 seconds, representing a considerable gain in imaging speed. Furthermore, the NN model reduced overall processing time for a single DOCT dataset from 4 hours to just 30 minutes, enabling significantly faster analysis and throughput. The team trained the NN model using a refinement strategy, building upon existing point-scanning OCT data to adapt it for use with FF-SS-OCM data, circumventing the need for extensive new dataset acquisition. Measurements confirm that the NN-generated LIV images are both qualitatively and quantitatively comparable to those derived from the conventional 32-volume method, maintaining image fidelity while drastically reducing computational load.

The breakthrough allows for high-resolution, label-free imaging of intratissue activity with extended imaging depth, making it particularly well-suited for in vitro sample analysis. This technology has the potential to accelerate drug development and large-scale screening processes by enabling high-throughput imaging of biological samples, potentially reducing the time required to analyze 96-well plates from over 10 hours to a more manageable timeframe. The study demonstrates a practical solution to the data bottleneck in FF-DOCT, paving the way for more efficient and rapid volumetric imaging of dynamic biological processes. Results demonstrate the feasibility of applying refined neural networks to complex optical coherence microscopy data, opening new avenues for real-time analysis and high-content screening applications.

Neural Networks Accelerate DOCT Imaging Significantly

Researchers have developed a novel neural network-based method to accelerate dynamic coherence tomography (DOCT) imaging using full-field swept-source coherence microscopy. This approach successfully generates high-definition images, specifically logarithmic intensity variance (LIV) images, from a significantly reduced number of input volumes, just four, compared to the conventional thirty-two. This refinement of an existing neural network model, achieved with a limited dataset of twenty samples, demonstrates its adaptability to full-field systems. The principal achievement lies in the substantial reduction of data handling requirements, decreasing data size, transfer time, and processing time by a factor of eight.

This improvement is particularly significant for full-field swept-source coherence microscopy, which inherently produces large datasets. The authors acknowledge a limitation in the current unavailability of the underlying data, though it may be accessible upon request. Future work could explore broader applications of this technique and expansion of the training dataset to further enhance the model’s robustness and generalizability, ultimately facilitating the practical implementation of this advanced imaging modality.