FIGURE 4. Deep learning classification of RI patches. (a) The architecture of the CNN classifier for RI patches. (b) The loss curve for the training set and the validation sets during 200 epochs of training. (c)-(d) validation and test confusion matrices.
3.3 Whole-slide Analysis of Thrombus Composition
To further assess the performance of the method in analyzing whole-slide images, the class of each patch was identified and stitched together to create a large field-of-view thrombus RI image. The patch classification results were used to create a virtually colored patch image and an unstained image, which were compared to the ground truth annotation map based on H&E-stained images. Figure 5 shows the results for four different regions, including slides 1, 2, and 3, and a region of interest (ROI) from slide 3. Despite the differences in resolution due to the patch-wise processing, the prediction based on DL and ODT showed a similar spatial distribution of RBC and fibrin regions as the ground truth. The sizes of the regions varied from 101 × 334 μm2 for the ROI to 2.71 × 5.58 mm2for slide 1.