Transfer Learning Strategies for Histopathology Patch Classification Using ResNet-18

Automated detection of tumor tissue in histopathology whole-slide images is a critical bottleneck in digital pathology. The PatchCamelyon (PCam) benchmark—327,680 lymph node patches labeled for metastatic tumor presence—provides a standardized testbed for this problem. In this work, we systematically compare four transfer learning strategies applied to ResNet-18 for binary patch classification on PCam: (S1) frozen backbone (feature extraction only), (S2) partial fine-tuning (early layers frozen, later layers trained), (S3) full fine-tuning of all pretrained weights, and (S4) training from scratch with random initialization. Each strategy is evaluated over three independent runs, reporting mean ± std of test accuracy, F1-score, and ROC-AUC alongside wall-clock training time per epoch. We additionally analyze convergence dynamics via learning curves, visualize first-layer convolutional filter evolution, and present GradCAM activation maps revealing how each strategy attends to tissue regions. S3 achieves the best test ROC-AUC of 0.9334 ± 0.0111, while S1 offers competitive performance (0.8514 ± 0.0017) at 42% lower training cost per epoch. We further demonstrate a confidence-based three-tier patch triage system enabling clinically meaningful prioritization.