Skip to main content
. 2022 Jul 20;4(5):e210299. doi: 10.1148/ryai.210299

Figure 1:

Flowchart of images used from different cohorts. (A) Split of dataset 1 (DS1), where five training sets containing different numbers of images and lesions were used for developing different versions of the models, a tuning set was used to select the best models, and a testing set was used for final evaluation. A subset was further randomly selected from the testing set for comparing the deep learning models with radiologists. (B) Frontal chest radiographs (CXRs) from the original CheXpert dataset were used as the additional training data. (C) Two subsets from the original National Institutes of Health (NIH) dataset were used for external testing. (D) The manually labeled posterior-anterior (PA) and anterior-posterior (AP) views from the PadChest dataset were used for external testing. LEs = lesions.

Flowchart of images used from different cohorts. (A) Split of dataset 1 (DS1), where five training sets containing different numbers of images and lesions were used for developing different versions of the models, a tuning set was used to select the best models, and a testing set was used for final evaluation. A subset was further randomly selected from the testing set for comparing the deep learning models with radiologists. (B) Frontal chest radiographs (CXRs) from the original CheXpert dataset were used as the additional training data. (C) Two subsets from the original National Institutes of Health (NIH) dataset were used for external testing. (D) The manually labeled posterior-anterior (PA) and anterior-posterior (AP) views from the PadChest dataset were used for external testing. LEs = lesions.