Automated Style-Aware Selection of Annotated Pre-Training Databases in Biomedical Imaging

Due to the broad use of deep learning and its need for big data, annotated and available databases for different tasks are constantly appearing. Nevertheless, they often remain unexploited due to the difficulty of effectively performing transfer learning between different databases. In medical imaging, the task of transfer learning is challenging due to: the variety of image modalities, organ/cell shapes, etc., and the lack of available and annotated data. In this paper, we propose an automated pipeline for predicting the similarity values of new database compared to known annotated databases. The system consists of an autoencoder trained on a comprehensive loss function that considers image reconstruction, style features, and dataset membership. A similarity measure is defined based on the resulting 2D latent space, which is demonstrated to have a correlation with the pre-training results on not annotated databases. Hence, our similarity measure could be used to select the most suitable known database for transfer learning or domain adaptation.