Background Transfer learning is widely used in medical imaging when labeled data are limited, but data augmentation policies are often selected empirically, and their interaction with partial fine-tuning is not always evaluated in a controlled manner. This study examined how data augmentation strength influences performance, seed-to-seed stability, and clinically relevant sensitivity-specificity trade-offs in binary pneumonia classification from chest X-ray images using a total of 5,856 chest X-ray images. Methodology Using a fixed ResNet-50 backbone, we compared linear probing with a frozen backbone, shallow fine-tuning of the last convolutional block, and deeper fine-tuning of the last two convolutional blocks under the following three augmentation regimes: no augmentation, light augmentation, and strong augmentation. Experiments were repeated across five random seeds and evaluated on a fixed held-out test set using accuracy, F1-score, sensitivity, specificity, area under the receiver operating characteristic curve, and area under the precision-recall curve. Results Across augmentation regimes, fine-tuning consistently outperformed linear probing, while differences between fine-tuning conv5 and fine-tuning conv4-5 were small. Relative to no augmentation, light augmentation improved accuracy, F1-score, and specificity while preserving high sensitivity. Strong augmentation yielded the highest mean accuracy and F1-score, but it did not consistently improve the area under the receiver operating characteristic curve or the area under the precision-recall curve relative to light augmentation. Fine-tuned models also showed lower variability across seeds than linear probing. Conclusions These findings indicate that augmentation strength is an important determinant of robustness and clinically relevant model behavior, and that augmentation policy should be selected jointly with fine-tuning depth in practical chest X-ray transfer learning pipelines.
Ackley Dias Will (Mon,) studied this question.