Figure 2.

Uniform Manifold Approximation and Projection (UMAP) visualization of neural network latent space and examples of false results. All input patches were fed into classification models and the corresponding post-convolution dense vectors (1024-length) were generated. All the vectors were mapped into 2-dimensional space using UMAP. Each point corresponds to a patch in (a) the tumor classifier, (b) the EBV classifier, and (c) the 3-Class classifier, colored by tissue type (tumor vs. normal) and tumor type (EBV positive vs. EBV negative). Histology images (d–f) of false predictions (red triangles) of each classifier (a–c), which are located apart from their corresponding cluster on UMAP. These points were examined to investigate potential reasons for misclassification.