From Adults to Children: Benchmarking Foundation Models for Chest X-Ray Classification under Domain Shift
This study tested how well foundation models generalize when classifying chest X-rays from adults to children. The results show that all models suffer significant performance drops, exceeding 10 percentage points, when moved from adult data to pediatric data. The findings indicate that the success of adaptation depends heavily on the model's initial training and the specific fine-tuning strategy chosen.