Adaptive Curved Slicing for En Face Imaging in Optical Coherence Tomography

Abstract

Optical coherence tomography (OCT) employs light to acquire high-resolution 3D images and is widely applied in fields such as ophthalmology and forensic science. A popular technique for visualizing the top view (en face) is to slice it with flat horizontal plane or apply statistical functions along the depth axis. However, when the target appears as a thin layer, strong reflections from other layers can interfere with the target, rendering the flat-plane approach ineffective. We apply Otsu-based thresholding to extract the object's foreground, then use least squares (with Tikhonov regularization) to fit a polynomial curve that describes the sample's structural morphology. The surface is then used to obtain the latent fingerprint image and its residues at different depths from a translucent tape, which cannot be analyzed using conventional en face OCT due to strong reflection from the diffusive surface, achieving FSIM of 0.7020 compared to traditional en face of 0.6445. The method is also compatible with other signal processing techniques, as demonstrated by a thermal-printed label ink thickness measurement confirmed by a microscopic image. Our approach empowers OCT to observe targets embedded in samples with arbitrary postures and morphology, and can be easily adapted to various optical imaging technologies.

Keywords: 3D metrology; functional sensing; image processing; optical coherence tomography; surface morphology fitting.

Figure A1

Average profile along the depth axis comparing curve slicing and flat slicing; sample according to multi-layer surface in Section 4.1.

Figure 1

Experimental setup: (a) equipment and samples, (b) translucent tape with latent fingerprint, and (c) a thermal-printed label.

Figure 2

Curved slicing method. (a) Original B-scan (cross-sectional) slice with threshold indicated as T. (b) Image after binarization (white voxels) and surface-fitting $F_1(\Delta z=0)$ (green line). (c) Inset indicating boundary score for automated boundary detection. (d) Otsu’s thresholding operation, with slice voxel histogram in background (left axis) and variances calculations in foreground (right axis). (e) 3D surface indicating curved en face slice with the gray sheet as a cross-sectional B-scan used in (a,b). (f) Flow chart describing our method.

Figure 3

Latent fingerprint residue detection. (a) Cross-section of translucent tape with bright yellow lines indicating tape boundaries. $F_2(\Delta z)$ indicates the curved slicing function (selecting the second layer). (b) Curved en face slice at $\Delta z=8$ showing another layer of residue. (c–e) Microscopic images at each region. (f) Inked reference from fingerprint. (g) Traditional average en face from $z=170$ to $z=435$. (h) Curved en face slice at $\Delta z=2$ showing a layer of resolvable latent fingerprint pattern. Images (b,f–h) are enhanced for clarity.

Figure 4

Ink thickness measurement in a thermal-printed label. (a) Cross-sectional microscopic image of a printed label paper showing ink thickness $L=21.0\mathsf{\mu }\mathrm{m}$. (b) Pattern signals at each $\Delta z$ slice. The $1/e^2$ of the Gaussian curve fit amplitude determines the layer count $l_z$, which converts to ink thickness $L=21.7\mathsf{\mu }\mathrm{m}$. (c) Three-dimensional curved slice at $\Delta z=-4$ showing a heart pattern with its projection (or flattening). Images are enhanced for clarity. (d) Unprocessed en face at $\Delta z=-4$; the middle square region is used to perform 2D Fourier transform. (e) A two-dimensional frequency-domain image with DC in the center. Arrows indicate a periodic printing pattern and high frequency noise. The subtraction of intensities results in a data point of the pattern signal on (b).

Figure 5

Automated ink thickness measurement results. (a) Intensity distribution sorted in descending order; each line represents each slice $\Delta z$, and the color of each line represents the ratio $P_2/P_{100}$ of each slice. (b) Ratio according to each $\Delta z$, where the maximum ratio corresponds to candidate slice $\Delta z_m$. The top inset shows dominant and recessive frequency-domain data points of $\Delta z_m$, where the points are aligned according to the inked pattern. The bottom inset shows frequency-domain data points of a non-candidate slice, where points are distributed in a random manner (with higher spatial frequency indicating possible noise). (c) Signal processing results of each case. (d) Box plot of measurement results, with standard deviations of $5.1802\mathsf{\mu }\mathrm{m}$ for different labels and $3.4322\mathsf{\mu }\mathrm{m}$ for different orientations.