Skip to main content
. 2023 Dec 13;5(1):100896. doi: 10.1016/j.patter.2023.100896

Figure 3.

Figure 3

Hierarchical attention network: Supervised cell-type classification

(A) Hierarchical attention network structure. Blue blocks: node-level neural networks. Yellow blocks: sentence-level neural networks. Orange blocks: neuron-level neural networks.

(B) The receiver operating characteristic (ROC) curve and area under the curve of the DSM-HAN classifier. Taking multi-class classification as 12 binary classifications, we calculated false positive rate and true positive rate for each binary classification and plot the ROC curves. The ROC curve of each type is displayed in one color.

(C) Comparison between methods. Each method was tested by 30 times of cross-validation, and results are displayed by boxplots (numbers above box: p values of the Mann-Whitney U rank test (one-side) on test accuracy between DSM-HAN and others).

(D) Robustness test: the testing accuracy for noise levels ranging from 10 to 1,000 μm. The accuracy curve is aggregated over repeated values (each noise level with 30 independent trainings and testings), showing the mean testing accuracy and 95% confidence interval.