Skip to main content
. 2022 Jun 3;10(6):1314. doi: 10.3390/biomedicines10061314

Figure 1.

Figure 1

Study framework. The fundus images were first preprocessed with data augmentation before being forwarded to the EfficientNet B3 algorithm to build the deep learning classifier. (A) Taipei Veterans General Hospital (TVGH) model was built using images from TVGH dataset and then fine-tuned as dataset-specific model using the weight of the TVGH model and new training images from specific external dataset to improve the diagnostic performance on the specific dataset. (B) Integrated model was built using combined images from TVGH dataset and all the external datasets.