XCT-SAM adapts Meta's SAM to industrial defect detection with 4.15M params

Defect segmentation in additive manufacturing X-ray computed tomography (XCT) images is notoriously difficult due to severe class imbalance and large distribution shifts across scan conditions. Although foundation models like Meta's Segment Anything Model (SAM) offer strong general-purpose segmentation priors, their natural-image pre-training transfers poorly to industrial XCT, where defects appear as subtle non-semantic microstructural anomalies. To address this, researchers from the University of Arkansas, Johns Hopkins University, and other institutions propose XCT-SAM, a sequential parameter-efficient adaptation framework that progressively bridges the domain gap.

The key innovation is a two-step adaptation process: first, Conv-LoRA adapters (with rank r=2) are fine-tuned on an alloy-microstructure dataset, injecting convolutional spatial inductive bias into SAM's backbone while training only about 4.15 million parameters—keeping over 99% of the model frozen. This intermediate domain adaptation prepares the model for the larger shift to real XCT data. Evaluated on out-of-distribution CycleGAN-XCT benchmarks and authentic NIST X-ray CT scans, XCT-SAM consistently outperforms zero-shot SAM and other domain-adapted SAM baselines, achieving the best overall Intersection over Union (IoU) and Dice scores. The method demonstrates that sequential adaptation with parameter-efficient fine-tuning is a practical, resource-efficient route for deploying large vision models in specialized industrial inspection tasks. Source code is publicly available on GitHub.