Impact of Transfer Learning and Model Capacity on Patch-Level Metastasis Detection Across Training Data Regimes

The development of deep learning models for histopathology frequently relies on transfer learning (TL) to overcome the scarcity of annotated medical data. However, the efficacy of this approach under varying degrees of data availability is still not fully understood. This study systematically evaluates the impact of TL and model capacity on patch-level breast cancer metastasis detection across different PatchCamelyon training data regimes (25%, 50%, and 100% of the 262,144 training patches) using three network configurations: a frozen ResNet18, a partially fine-tuned ResNet18, and a custom small convolutional neural network (SmallCNN). The results demonstrate that while the partially fine-tuned ResNet18 achieved the highest performance under limited-data conditions (Area Under the Precision-Recall Curve (AUPRC) 0.9020 at 25% data), the compact SmallCNN trained from scratch outperformed all pre-trained variants when the full dataset was available (AUPRC 0.9279). These findings indicate that while TL is highly beneficial for data-scarce scenarios, compact domain-specific architectures can achieve superior predictive performance given sufficient task-specific data.