Neuronal feature embedding and classification in the recognition model.
(A) Distance matrix between neuron clusters in the feature embedding space. The matrix represents the cosine distance between trained weight vectors in the final layer of the recognition neural network (Fig 6C and 6D). wi can well approximate the direction towards the corresponding neuron cluster center. (S2(B) and S2(C) Fig). The matrix indices, representing digital IDs, are rearranged by the hierarchical clustering method. Dmin = 0.28, Dmax = 1.62 when values along the diagonal are excluded, suggesting large inter-cluster distance. (B) t-SNE visualization of the feature embedding space. The low-dimensional representation reveals intra-class similarity (S2(A) Fig) and inter-class differences among neurons (related to A). For the sake of clarity, feature vectors from 4% of the training, validation and test volumes are drawn. (C) Train CeNDeR on C1 with different number of randomly chosen volumes and then test on the remaining volumes of C1. The top-1 accuracy reaches 88.25% with 30 training volumes (star), while the triangle denotes a 95.48% accuracy with 130 training volumes. (D) Overall tracking accuracy of C1 across 200 test volumes when the neural network was trained on 130 volumes. Inset shows MIP images of two cases with low recognition accuracy. (E) Train CeNDeR on NeRVE dataset with different number of randomly chosen volumes and then test on the remaining volumes of NeRVE. The top-1 accuracy reaches 84.10% (star) with 80 training volumes, the best performance among existing tracking methods (see Table 4). The triangle denotes a 86.51% accuracy with 130 training volumes. (F) Overall tracking accuracy of NeRVE dataset across 1222 test volumes when the network was trained with 130 volumes. In D,F, the horizontal axis represents the volume index over time.