Reconstruction of Optical Coherence Tomography Images from Wavelength Space Using Deep Learning

Abstract

Conventional Fourier domain Optical Coherence Tomography (FD-OCT) systems depend on resampling into a wavenumber (k) domain to extract the depth profile. This either necessitates additional hardware resources or amplifies the existing computational complexity. Moreover, the OCT images also suffer from speckle noise, due to systemic reliance on low-coherence interferometry. We propose a streamlined and computationally efficient approach based on Deep Learning (DL) which enables reconstructing speckle-reduced OCT images directly from the wavelength (λ) domain. For reconstruction, two encoder-decoder styled networks, namely Spatial Domain Convolution Neural Network (SD-CNN) and Fourier Domain CNN (FD-CNN), are used sequentially. The SD-CNN exploits the highly degraded images obtained by Fourier transforming the (λ) domain fringes to reconstruct the deteriorated morphological structures along with suppression of unwanted noise. The FD-CNN leverages this output to enhance the image quality further by optimization in the Fourier domain (FD). We quantitatively and visually demonstrate the efficacy of the method in obtaining high-quality OCT images. Furthermore, we illustrate the computational complexity reduction by harnessing the power of DL models. We believe that this work lays the framework for further innovations in the realm of OCT image reconstruction.

Keywords: image reconstruction; optical coherence tomography; speckle noise; time complexity.

Figure 1

Schematics of the proposed framework containing Spatial Domain–Convolution Neural Network (SD-CNN) and Fourier Domain–Convolution Neural Network (FD-CNN); $x_i$ and $x_i^{\prime }$ are the outputs and $y_i$ and $y_i^{\prime }$ are the ground truths for SD-CNN and FD-CNN, respectively.

Figure 2

(a) DL network used as the Fourier Domain–Convolution Neural Network and Spatial Domain–Convolution Neural Network; (b) 1 B-scan and image averaging using 5, 7, 9 B-scans for vein, lemon, and cherry; (c) Wavenumber layer interleaved with the input of Spatial Domain–Convolution Neural Network.

Figure 3

Comparison of k-linearization fringes for different volumes used: vein, finger, lemon, tooth, cherry, flounder-egg, seed (pea).

Figure 4

Comparison of B-scans from five different volumes. (A) Ground truth (seven B-scans of OCT system output averaged), (B) OCT system output, (C) OCT system raw data input, (D) output of the proposed framework. The reconstruction (D) shows high similarity to the desired ground truth.

Figure 5

Comparison of line plot to show the variation in intensity (A.U.) for the central column of the red rectangle marked in Figure 4; here, plots correspond to (a) vein, (b) finger (c) lemon, (d) tooth, and (e) cherry samples, for ground truth, input, OCT output, and proposed framework outputs shown using different lines in each plot.

Figure 6

Comparison between the ground truth, the output of the Spatial Domain–Convolution Neural Network and Fourier Domain–Convolution Neural Network for (a) vein, (b) finger, (c) lemon, (d) tooth, and (e) cherry. For each image, the magnified regions are shown for better comparison. The results of the combined SD-CNN+FD-CNN show enhanced performance when compared to output of only SD-CNN, demonstrating the better reconstruction capability of high-frequency details using FD-CNN.

Figure 7

Comparison of cross-validation results on (a,c,e) flounder egg and (b,d) seed (pea) samples for ground truth, OCT output, and reconstructions obtained by the proposed model. The cross-validation results show robustness and generalization capability of the proposed model on a completely unseen volume.