Neural-network-based high-speed and high-definition full-field dynamic optical coherence tomography

TL;DR

This work tackles the data bottleneck in high-definition full-field swept-source DOCT by developing a neural-network that generates LIV-based DOCT images from just four volumes. By refining a basement model trained on lower-resolution, point-scanning data with a small FF-SS-OCM dataset, the authors enable LIV reconstruction from four FF-SS-OCM volumes with fidelity comparable to the ground truth computed from 32 volumes. The four-volume protocol reduces data size by about eightfold and slashes data transfer and processing times from 7 minutes/4 hours to roughly 55 seconds/30 minutes, respectively, demonstrated on breast cancer spheroids with partial generalization to HT-29 spheroids. While promising for high-throughput FF-DOCT, the method remains specific to LIV-based contrast and spheroidal samples, and extensions to other time-order-sensitive DOCT modalities may require additional refinements.

Abstract

A neural-network (NN)-based method for high-speed, high-definition dynamic optical coherence tomography (DOCT) using full-field swept-source optical coherence microscopy (FF-SS-OCM) is demonstrated. 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. We address this issue by introducing an NN-based DOCT method that generates high-definition logarithmic intensity variance (LIV) -based DOCT images from only four OCT volumes, whereas the conventional method required 32 volumes. The NN model successfully generates an LIV image that is qualitatively and quantitatively similar to the LIV image computed from 32 volumes. This approach significantly reduces data size, transfer time, and processing time for DOCT imaging by a factor of eight. Specifically, these were reduced from 42 GB to 5.3 GB, 7 min to 55 s, and 4 hours to 30 min, respectively.

Figures (8)

• Figure 1: NN architecture used for LIV generation. The NN consists of three parts: an encoder (red dashed-line box), a decoder (blue dashed-line box), and a skip connection (brown arrows). The input to the NN is a set of four cross-sectional OCT images, which were acquired at the same position in the sample but at different time points, and the output is an LIV image, which is a DOCT image defined as the time variance of the dB-scaled OCT images. The leftmost part indicates the spatial dimensions of the image at each step of the U-Net, where (H, W) represents the original size of the input image.
• Figure 2: Schematic diagram of the model refinement. The basement model has been trained by low-resolution point-scanning OCT images. Refinement is performed with FF-SS-OCM data sequences of only 20 spheroids, and finally gives the refined NN model.
• Figure 3: Schematic diagram of two types of volume acquisition protocols. (a) 32-volume acquisition protocol, where the inter-volume time separation is 0.5 ms and the total acquisition time is 8.02 s. The four volumes with orange color are extracted from the 32 volumes as input to the NN. (b) Four-volume acquisition protocol, where the time separation is 2.15 s and the total acquisition time is 7.45 s. The inter-volume time separation of 2.15 s is identical to that of the four volumes extracted from the 32 volumes.
• Figure 4: Comparison of the GT, refined LIV, and non-refined LIV images of human breast cancer spheroids. The left and right halves of the figure show the en-face and cross-sectional images, respectively. The depth positions of the en-face images are indicated by black arrows on the right side of the cross-sectional images. The last two rows are magnified views of representative cases. The scale bar represents 100 $\muup$m. The output of the non-refined NN model exhibits lower LIV values than the GT and artifactual horizontal elongation patterns.
• Figure 5: Comparison of the GT, refined LIV, and non-refined LIV images. Note that the GT images are generated from 32-volume sequence and are identical to those in Fig. \ref{['fig:modelEvaluation']}, while the refined and non-refined images were generated from the dataset obtained by the four-volume acquisition protocol. The left and right halves of the figure show the en-face and cross-sectional images, respectively. The depth positions of the en-face images are indicated by black arrows on the right side of the cross-sectional images. The scale bar represents 100 $\muup$m. The refined images are similar to the GT.