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

2026-07-15 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition · Depth: Expert, quick

Summary

XCT-SAM introduces a sequential parameter-efficient adaptation framework designed for industrial X-ray computed tomography (XCT) defect segmentation in additive manufacturing (AM). This addresses challenges like severe class imbalance and large distribution shifts, which cause foundation models like the Segment Anything Model (SAM) to perform poorly due to their natural-image pre-training. Instead of direct adaptation, XCT-SAM first fine-tunes Conv-LoRA adapters on an alloy-microstructure dataset, then transfers the adapted model to XCT images, progressively bridging the domain gap. Using Conv-LoRA with rank r=2, the framework trains approximately 4.15M parameters while keeping over 99% of SAM frozen. Evaluated on CycleGAN-XCT benchmarks and real-world NIST XCT scans, XCT-SAM consistently outperforms zero-shot SAM and other domain-adapted baselines, achieving superior IoU and Dice scores.

Key takeaway

For Machine Learning Engineers adapting foundation models like SAM to specialized industrial domains with limited labeled data, XCT-SAM's sequential, parameter-efficient approach offers a robust strategy. You should explore multi-stage domain adaptation with Conv-LoRA to inject spatial inductive bias and progressively bridge large domain gaps, significantly improving performance on challenging industrial XCT defect segmentation tasks.

Key insights

Sequential parameter-efficient domain adaptation effectively bridges large domain gaps for specialized image segmentation.

Principles

• Intermediate domain adaptation improves transfer learning.
• Parameter-efficient adapters inject convolutional spatial inductive bias.
• Low-rank Conv-LoRA (r=2) is effective for vision model adaptation.

Method

Fine-tune Conv-LoRA adapters on an intermediate alloy-microstructure dataset, then transfer the adapted model to target XCT images to progressively bridge the domain gap.

In practice

• Employ Conv-LoRA for efficient domain adaptation in vision.
• Implement multi-stage fine-tuning for large domain shifts.
• Consider low-rank adapters (r=2) to minimize trainable parameters.

Topics

• XCT-SAM
• Defect Segmentation
• Domain Adaptation
• Parameter-Efficient Fine-Tuning
• Segment Anything Model
• Additive Manufacturing

Code references

• Mahedi-61/XCT-SAM

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• SynSur: An end-to-end generative pipeline for synthetic industrial surface defect generation and detection — Synthetic data, when augmenting scarce real datasets, can improve industrial defect detection.
• SAM3 Self-Distillation for Fine-Grained GOOSE 2D Semantic Segmentation — Self-distillation using SAM3 as a teacher, combined with image-level multi-scale augmentation and aggressive photometric distortion, significantly boosts fine-grained semantic segmentation.
• SPDCN: Strip-based Deformable Convolutional Network for Steel Surface Defect Segmentation — SPDCN uses strip-based deformable convolutions and fuzzy attention to segment anisotropic steel defects, outperforming existing methods.
• Vision Transformer-Conditioned UNet for Domain-Adaptive Semantic Segmentation — ViTC-UNet combines frozen ViT global priors with UNet local inductive biases for efficient, high-precision biomedical semantic segmentation.
• Training-free Cross-domain Few-shot Segmentation via Robust Semantic Representation and Matching — A training-free framework built on DINOv3 achieves state-of-the-art CD-FSS by enhancing semantic representation and matching without overfitting.
• Domain-Guided Prompting of the Segment Anything Model for Seismic Interpretation: The Role of Attributes, Visualization, and Hybrid Prompts — Zero-shot SAM adaptation for seismic interpretation is achieved by combining domain-guided visualization with hybrid prompting.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.