Explainability Methods for AI-Assisted Diagnosis of Lymph Node Metastases in Digital Pathology: A Quantitative Comparative Study

Background/Objectives: Artificial intelligence (AI) systems for detecting lymph node metastases in histopathological images achieve near-expert classification performance but remain opaque to clinicians, limiting their clinical adoption and regulatory acceptance. This study presents the first rigorous quantitative framework for evaluating and comparing explainable AI (XAI) methods in digital pathology, providing actionable evidence-based guidance for clinical deployment. Methods: Four XAI techniques—LIME, GradCAM, GradCAM++, and SHapley Additive exPlanations (SHAP) via DeepExplainer—were applied to three convolutional neural networks (VGG19, ResNet50, and EfficientNetB3) trained on the PatchCamelyon (PCam) benchmark (220,026 patches). Quantitative evaluation employed two complementary frameworks: spatial agreement with expert pathologist annotations (Intersection over Union and Sørensen–Dice coefficient on 2847 annotated patches) and faithfulness metrics (Area Over the Perturbation Curve and insertion/deletion Area Under the Curve) independent of external annotations. Threshold sensitivity analysis was also conducted at fixed binarisation thresholds (τ = 0.3 and τ = 0.7) in addition to Otsu automatic thresholding. Results: GradCAM++ achieved the highest spatial agreement with pathologist annotations (mean IoU = 0.52 ± 0.14 for EfficientNetB3), while SHAP yielded the highest faithfulness scores (AOPC = 0.61 ± 0.08). The parameter-free squaregrid LIME variant offered a favourable trade-off (IoU = 0.44 ± 0.17) at 3.8× lower computational cost than LIME AVG. Relative method rankings were preserved across all binarisation thresholds, confirming the robustness of the evaluation framework. A Spearman correlation of ρ = 0.81 was found between model classification AUC and spatial agreement, indicating that superior classification performance systematically produces more spatially coherent explanations. Conclusions: GradCAM++ is recommended for high-throughput clinical workflows; SHAP for research contexts requiring maximal faithfulness; and squaregrid LIME as a transparent, parameter-free baseline for clinical communication and audit, preferred over LIME AVG on account of its parameter-free operation and 3.8× lower computational cost. A tiered deployment strategy integrating GradCAM++, SHAP, and squaregrid LIME is proposed. These findings provide quantitative, technical evidence of a type relevant to regulatory frameworks such as the FDA SaMD Action Plan and EU IVDR 2017/746; formal regulatory acceptance would additionally require prospective, multi-site external validation and a pathologist reader study, which lie beyond the scope of this single-benchmark study.