[Paper Review] Automatic Volumetric Segmentation of Additive Manufacturing Defects with 3D U-Net

Vivian Wen Hui Wong, Max Ferguson+3 more|arXiv (Cornell University)|Jan 22, 2021

Additive Manufacturing and 3D Printing Technologies29 references22 citations

TL;DR

This paper proposes a 3D U-Net deep learning model for automatic volumetric segmentation of defects in X-ray computed tomography (XCT) scans of additively manufactured parts. By adapting medical imaging techniques to additive manufacturing, the method achieves a mean intersection over union (IoU) of 88.4%, demonstrating the first successful application of 3D U-Net for defect segmentation in AM with high accuracy on challenging low-contrast, small defects.

ABSTRACT

Segmentation of additive manufacturing (AM) defects in X-ray Computed Tomography (XCT) images is challenging, due to the poor contrast, small sizes and variation in appearance of defects. Automatic segmentation can, however, provide quality control for additive manufacturing. Over recent years, three-dimensional convolutional neural networks (3D CNNs) have performed well in the volumetric segmentation of medical images. In this work, we leverage techniques from the medical imaging domain and propose training a 3D U-Net model to automatically segment defects in XCT images of AM samples. This work not only contributes to the use of machine learning for AM defect detection but also demonstrates for the first time 3D volumetric segmentation in AM. We train and test with three variants of the 3D U-Net on an AM dataset, achieving a mean intersection of union (IOU) value of 88.4%.

Motivation & Objective

• To address the challenge of automatic defect segmentation in additive manufacturing due to low contrast, small size, and variable appearance in XCT images.
• To adapt 3D convolutional neural networks, specifically the 3D U-Net architecture, from medical imaging to the domain of additive manufacturing.
• To evaluate the performance of multiple 3D U-Net variants on a real AM defect dataset for volumetric segmentation.
• To establish a benchmark for deep learning-based 3D segmentation of defects in AM components using XCT data.
• To enable automated, reliable quality control in additive manufacturing through precise defect localization in 3D.

Proposed method

• A 3D U-Net architecture is trained end-to-end on 3D volumetric XCT data of additively manufactured samples to segment defects.
• The model uses encoder–decoder architecture with skip connections to preserve spatial resolution and improve feature learning across scales.
• Three variants of the 3D U-Net are trained and evaluated on a custom AM defect dataset to assess performance differences.
• Data augmentation techniques are applied to improve generalization, especially given the limited size of defect datasets.
• The network is trained using binary cross-entropy loss with sigmoid activation to predict pixel-wise defect probabilities.
• Inference is performed on 3D volumes to generate dense segmentation masks identifying defect regions in 3D space.

Experimental results

Research questions

• RQ1 Can a 3D U-Net model effectively segment small, low-contrast defects in XCT scans of additively manufactured parts?
• RQ2 How does the performance of 3D U-Net compare across different architectural variants on AM defect segmentation?
• RQ3 To what extent can transfer learning techniques from medical imaging be adapted to improve defect segmentation in additive manufacturing?
• RQ4 What is the achievable mean intersection over union (IoU) for 3D volumetric segmentation of AM defects using deep learning?
• RQ5 Can automatic 3D segmentation reduce the need for manual inspection in AM quality control pipelines?

Key findings

• The proposed 3D U-Net model achieved a mean intersection over union (IoU) of 88.4% on the AM defect dataset, indicating high segmentation accuracy.
• Among the three 3D U-Net variants tested, one achieved the highest IoU, demonstrating the importance of architectural design in defect segmentation.
• The model successfully segmented defects with poor contrast and small spatial extent, which are common challenges in XCT-based AM inspection.
• The results validate the transferability of 3D U-Net from medical imaging to additive manufacturing, marking a first in this domain.
• The method enables fully automatic, 3D volumetric segmentation, reducing reliance on manual defect annotation.
• The study establishes a new benchmark for deep learning-based 3D segmentation of defects in AM using XCT data.