Fig. 2

a presents the architecture of the deep neural network (DNN). This ensemble learning framework relies on pretraining two DenseNet models separately on two large public datasets (MIMIC and CheXpert) and fine-tuning them on our in-house mycobacterial datasets. b illustrates the developing process of our DNN and the data flow. The validation set was used to select the best DNN models. Patients in the internal test set came from the same data distribution as the internal training set. Patients in the external test set came from a different data distribution, bringing more challenges to the model