Hy-SAN: a hybrid shape-aware network for artifact-suppressed sparse-view computed tomography of through-silicon vias

High-throughput and non-destructive inspection of Through-Silicon Vias (TSVs) is a critical task in advanced semiconductor packaging. Sparse-view X-ray Computed Tomography (XCT) can accelerate TSV inspection but suffers from strong artefacts in analytical reconstructions and structural distortions in domain-agnostic Deep Learning (DL) models. We address these challenges with a hybrid image-domain refinement framework that first performs analytical reconstruction and then applies DL-based refinement. We introduce the Hybrid Shape-Aware Network (Hy-SAN) for the refinement stage. At the core of Hy-SAN is a Shape-Aware Attention (SAA) module, which learns and leverages the specific morphological priors of TSVs to guide the refinement process. Complemented by an efficient Hybrid Lightweight Block (HLB) and a Multi-scale Prior Injection Module (MPIM), Hy-SAN removes artefacts while preserving crucial structural details. On a TSV dataset, Hy-SAN achieves state-of-the-art performance, reaching an average PSNR of 39.18 dB for reconstructed images with a compact model of 6.10 M parameters. Hy-SAN provides an efficient solution for artefact-suppressed reconstruction in sparse-view XCT, achieving the fastest inference among all compared DL methods and enabling reliable downstream defect detection.