An Empirical Study of Dataset Size Effects on Fine-Tuning Depth in Transfer Learning for Chest X-ray Classification

Transfer learning is a standard approach in medical imaging to address the limited availability of annotated data. However, practical decisions regarding fine-tuning depth are often made heuristically, without clear empirical guidance on how these choices depend on dataset size. This study presents a controlled empirical analysis of the interaction between dataset size and fine-tuning depth in a chest X-ray image classification task using a total of 5856 chest X-ray images. Using a fixed ResNet-50 backbone, we compared linear probing, shallow fine-tuning of the last convolutional block, and deeper fine-tuning of the last two convolutional blocks across multiple data regimes. Linear probing remains competitive in low-data settings. As dataset size increases, performance improves across all strategies, and fine-tuning provides small but consistent gains in ranking-based metrics, including the area under the receiver operating characteristic curve (ROC-AUC) and average precision (AP), relative to linear probing. The gap between shallow and deeper fine-tuning narrows in larger regimes, indicating diminishing returns from additional unfreezing under this experimental setup. Rather than proposing a new model, this work provides practical guidance for selecting fine-tuning strategies based on data availability and system constraints, supporting informed engineering decisions in medical imaging pipelines where trade-offs between performance, complexity, and computational cost must be balanced.