Abstract
Purpose :
This study is to design and validate a fully convolutional network (FCN) architecture to compensate for system dispersion in optical coherence tomography (OCT).
Methods :
Our proposed FCN method is based on a modified UNet architecture (Fig. 1A) that uses an encoder-decoder pipeline. The input of the deep leaning pipeline can be single or multiple-channel OCT B-scans. Each B-scan was compensated by different second-order dispersion compensation coefficients to optimize different layers sequentially. The output obtained from the network was a fully compensated OCT B-scan where the dispersion was compensated at all depths. A lab-built SD-OCT system with 840 nm central wavelength was used for the experimental imaging. The axial and lateral pixel resolutions were 1.5 µm and 5 µm, respectively. Nine (7 train and 2 test) different OCT volumes with a field of view 3.5 mm x 3.5 mm were acquired from healthy human subjects. The proposed method was trained and tested using 1, 3, 5, 7, and 9 input channel models. Quantitative analysis was performed using peak signal to noise ratio (PSNR) and structural similarity index metric calculated at multiple scales (MS-SSIM).
Results :
High quality all depth compensated OCT B-scans were obtained when the proposed model was implemented using 5, 7, and 9 input channels. Output from the 1 and 3 input channel models also demonstrated better resolution retaining more structural information than a raw uncompensated image. High similarity between the output images and the ground truth was observed and the MS-SSIM score for 5, 7, and 9 input channel models were 0.97 ± 0.016, 0.97 ± 0.014, and 0.97 ± 0.014, respectively. High signal strength compared to the background noise was also observed for these models with PSNR values of 29.95 ± 2.52, 29.91 ± 2.13, and 29.64 ± 2.26 dB for 5,7 and 9 input channel models respectively. Fig. 1B illustrates the ground truth (B1), uncompensated (B2), and output B-scan from 5 input channel model (B3).
Conclusions :
The proposed FCN can compensate for the dispersion at all depths and thus provides a feasible solution for fully automated dispersion compensation in OCT.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.