Inter vs. Intra Domain Study of COVID Chest X-Ray Classification with Imbalanced Datasets

Alejandro Galán-Cuenca, Miguel Mirón, Antonio Javier Gallego, Marcelo Saval-Calvo, Antonio Pertusa

Abstract: Medical image classification datasets usually have a limited availability of annotated data, and pathological samples are usually much scarcer than healthy cases. Furthermore, data is often collected from different sources with different acquisition devices and population characteristics, making the trained models highly dependent on the data domain and thus preventing generalization. In this work, we propose to address these issues by combining transfer learning, data augmentation, a weighted loss function to balance the data, and domain adaptation. We evaluate the proposed approach on different chest X-Ray datasets labeled with COVID positive and negative diagnoses, yielding an average improvement of 15.3% in $${\text {F}}_1$$ compared to the base case of training the model without considering these techniques. A 19.1% improvement is obtained in the intra-domain evaluation and a 7.7% for the inter-domain case.