Connected vs. Isolated Pore Classification via Learned Refinement of Connected-Component Analysis

Abstract The classification of isolated versus interconnected pores in porous media is critical for structural and fatigue assessments, yet classical 26-connectivity connected-component (CC) labeling fails catastrophically under realistic computed tomography (CT) noise due to the formation of false artifact webs. This study proposes a two-stage machine learning refinement pipeline to correct classical thresholding errors. Following Watershed over-segmentation to generate a proposal graph, we extract handcrafted geometric features—specifically minimum neck radius and path intensity—which successfully separate physical bridges from disconnected noise. We evaluate an XGBoost classifier, a GraphSAGE Graph Neural Network (GNN), and a 3D-CNN against the unrefined CC baseline. Feature importance analysis confirms that the tree-based model anchors its predictions on true physical geometry rather than random noise. While Watershed over-segmentation causes localized false-split penalties under the global Adjusted Rand Index (ARI), the learned refinement models achieve robust edge-level connectivity classification (Validation F1 ≥ 0.94), successfully severing the noise-induced bridge artifacts that compromise classical 26-connectivity thresholding.