FIGURE 3. Optical measurement. (a) BF image and (b) RI focal section. The boxed regions i, ii and iii denote mixed, RBC, and fibrin regions respectively. The size of each patch is 10.88 µm × 10.88 µm. Scale bars = 500 µm (boxed regions) and 50 µm (patches).
To validate the quality of the RI reconstruction, we visually compared the thrombus structure in the paired RI tomogram section and BF image. Figure 3 shows the BF and RI images of thrombus sample 3 at various magnifications: whole-slide level, multiple patches level, and individual patch level. The RI tomograms of the thrombus show various RI value distributions ranging from 1.48 to 1.5. Compared to the BF image, the RI tomogram section visualizes RBCs and fibrin structures of thrombus components better. The contrast of RI is directly related to the concentration of material.[36,37] In Figure 3, the magnified region of interest (ROI) of sample 3 is presented. In the RI tomogram section (Figure 3b), some borders of individual RBC are clearly shown even in tightly compressed region, whereas in the corresponding BF image (Figure 3a), the borders of individual RBCs are unclear. Some RBC borders in the RI image possibly indicate polyhedrocytes, which are tightly packed red blood cells (RBCs) with polyhedral shapes formed when blood clots contract. In the fibrin region, (iii) of Figure 3, the pores in the meshwork of fibrin fibers are also better observed in the RI tomogram. By contrast, the boundary of fibrin pores cannot be identified in the BF image.
3.2 Patch-wise Composition Prediction based on DL
Our dataset for DL-based prediction consisted of 166,887 patches of focal RI sections, with a FOV of each patch at 10.88 μm × 10.88 μm. For each slide, the patches were randomly separated into the training, validation, and test sets in a ratio of 7:1.5:1.5.
The DL classifier was trained using the training and validation datasets and blind-tested using the test dataset (Figure 4a ). The optimal model was chosen based on the validation loss during DL training, and the learning curve is presented in Figure 4b. The resulting test accuracy of the DL model was 95.1% (Figure 4d). A major error in the patch-wise classification was the misclassification of RBC into fibrin, accounting for 10.4% of all the RBC patches in the test set. This error could be due to the limited correspondence between the two consecutive thrombus sections, resulting in imperfectly mixed regions in the registered patch annotation. Additionally, the rate of confusing fibrin with the background was higher (2.97%) than that of confusing RBC with background (0.577%). This difference in background-related error is consistent with our observations of thin fibrin structures at the edge of the slide and the relatively low RI distribution of fibrin.