A 2D Sinogram-Based Approach to Defect Localization in Computed Tomography

TL;DR

This work tackles defect localization in CT by operating in sinogram space rather than the reconstructed image domain. It introduces a three-step deep learning pipeline—sinogram segmentation with a U-Net, instance segmentation via skeletonization and reclassification, and defect analysis from projection geometry—to detect and localize defects directly in sinograms, achieving a mean IoU of $0.9202$ and a defect position error of $1.3$ pixels on a $512$-detector. The approach demonstrates strong performance on simulated 2D parallel-beam data (MagicCube) and, to a lesser extent, on KiTS23 medical volumes, suggesting potential for real-time defect detection and integration with reconstruction workflows. Future work aims to extend the method to 3D and cone-beam geometries, moving beyond simulated projections toward practical CT applications.

Abstract

The rise of deep learning has introduced a transformative era in the field of image processing, particularly in the context of computed tomography. Deep learning has made a significant contribution to the field of industrial Computed Tomography. However, many defect detection algorithms are applied directly to the reconstructed domain, often disregarding the raw sensor data. This paper shifts the focus to the use of sinograms. Within this framework, we present a comprehensive three-step deep learning algorithm, designed to identify and analyze defects within objects without resorting to image reconstruction. These three steps are defect segmentation, mask isolation, and defect analysis. We use a U-Net-based architecture for defect segmentation. Our method achieves the Intersection over Union of 92.02% on our simulated data, with an average position error of 1.3 pixels for defect detection on a 512-pixel-wide detector.

Figures (5)

• Figure 1: (a) Input of the U-Net, pixels containing information of defects are painted red; (b) Result of sinogram segmentation is a binary mask that contains information about all found defects; (c) Result of instance segmentation, the information in (b) is separated based on which defect it belongs to.
• Figure 2: (a) Result of semantic segmentation, which is also the input of the instance segmentation algorithm; (b) Raw result of skeletonization. After erasing the intersection part we get the final skeletonization result (c). (d) is the result of reclassification. For ease of presentation, the results of the skeletonization in (b), (c), and (d) have been widened.
• Figure 3: Appropriate strategies vary with defect shape. (a) For circular defects, the CircleBox method outperforms the overlap method of two different projections in terms of accuracy. (b) For non-circular defects, the CircleBox approach is inappropriate.
• Figure 4: (a) Sample from the custom dataset. (b) Sample from KiTS23 kidney dataset. (c) Sample of a defect used in this paper which has a radius of 10 pixels.
• Figure 5: Our method can detect and analyze the defects before reconstruction and facilitate defect identification in reconstructed domain.