XCT-SAM: Sequential Parameter-Efficient Domain Adaptation of SAM for Industrial XCT Defect Segmentation

Additive manufacturing (AM) produces complex parts, but internal defects like pores and inclusions are difficult to detect due to their microscopic size and sparse distribution. While foundation models like the Segment Anything Model (SAM) are powerful, they are pre-trained on natural images and struggle to generalize to the low-contrast, non-semantic nature of XCT scans. This paper investigates how to efficiently adapt SAM to the AM XCT domain while overcoming the scarcity of labeled data and the significant domain shift.

The authors propose XCT-SAM, a framework that uses parameter-efficient fine-tuning (PEFT) to adapt SAM. Instead of a direct transfer from natural images to XCT data, the authors implement a two-stage curriculum: first, they fine-tune Conv-LoRA adapters on an alloy-microstructure dataset to align the model with metallic textures, and then they transfer these weights to the target AM XCT defect data. By using Conv-LoRA with a low rank (r=2), the framework injects spatial inductive bias into the frozen SAM backbone, training only 4.15M parameters (less than 1% of the total model). The authors also employ a Dice-Focal loss function to address the extreme class imbalance where defects occupy a tiny fraction of the total image area.

XCT-SAM consistently outperforms zero-shot SAM and other domain-adapted baselines across both synthetic (CycleGAN-generated) and real-world (NIST) XCT benchmarks. The two-stage adaptation strategy provides a more stable optimization path than direct fine-tuning, leading to higher IoU and Dice scores. The results demonstrate that intermediate domain adaptation is a highly effective strategy for specializing large foundation models for industrial inspection tasks where labeled data is limited.

This work provides a scalable path for deploying foundation models in industrial quality assurance. By reducing the need for massive, manually labeled XCT datasets and minimizing the computational cost of fine-tuning, XCT-SAM enables more accurate and automated detection of critical internal defects, which is essential for the certification and structural reliability of high-value AM components.