2 METHODS
2.1 Sample Preparation
Thrombi were retrieved from the middle cerebral artery by stent-based thrombectomy using a Solitaire stent (Covidien, Irvine, California) from a patient with AIS in Soonchunhyang University, Bucheon Hospital. The thrombi were sliced into sections (4-μm thickness) using a microtome. Two consecutive sections of a thrombus were visualized using two different imaging modalities: ODT and bright-field microscopy. Although the two consecutive sections may differ in microstructure, we assumed that they had nearly identical compositions of red blood cells (RBCs) and fibrin. One thrombus section was prepared with H&E staining, while the other was prepared without staining. Both thrombus sections were mounted on standard glass slides and sealed with coverslips (Figure 2 ).
2.2 Optical Measurement
The paired H&E stained and unstained slides of the thrombus were imaged with a BF slide scanner (Axio Scan.Z1; Zeiss; Jena, Germany) and ODT, respectively (Figure 2b). An ODT system was based on Mach–Zehnder interferometry and equipped with a digital micromirror device (DMD) for high-speed angle scanning (Figure S1 ). A blue continuous-wave laser (λ0 = 457 nm, Cobolt Twist, Cobolt) was split into a sample and reference beam using a beam splitter. The sample beam was modulated by the DMD for illumination angle control by impinging on a sample slide through a condenser objective (LUCPLFLN40X, NA = 0.6, Olympus). Then, the scattered fields from the sample at each illumination angle were collected by an objective lens (UPlanSAPO20X, NA = 0.75, Olympus) and interfered with the reference beam at the camera plane, generating a spatially modulated interferogram. Using field retrieval based on spectral filtering, complex field information consisting of amplitudes and phases was retrieved (Figure S1). The 2D optical field information at various illumination angles was used to reconstruct a 3D refractive index (RI) tomogram based on Fourier diffraction theorem.[29,30] Our setup had a lateral resolution of 0.17 μm and an axial resolution of 1.4 μm based on the Lauer criteria.[31,32]
To obtain whole-slide RI tomograms, multiple tomograms were stitched following the measurements in separate locations.[33] A phase-correlation algorithm was used to determine the relative position between overlapping tomogram tiles, and stitch them into tomograms.[33,34] By automatically stitching the scanned tomograms, the ODT setup achieved a millimeter-scale FOV (10 mm2). The stitched tomogram had a size of approximately 104 voxels in both lateral directions and 102 voxels in the vertical direction. The BF images acquired using a slide scanner were 2D images. Hence, the 3D RI tomograms were converted to 2D by identifying the optimal focal plane. For training, RI values between 1.48 and 1.5 were normalized to a grayscale from 0 to 255.
We registered the RI and BF images to annotate each RI patch based on the corresponding BF patch. Subsequently, we adjusted the stained bright-field image to match the size and FOV of the unstained QPI. Rescale, rotation, skew, and distortions were applied to the bright-field images using Adobe Photoshop (Version 2019, Adobe Systems, San Jose, CA, USA).