AI-Based Prediction of Total Porosity Volume Fraction from 3D CT Volumes: A Volumetric Regression Approach
Computed Tomography (CT) is a premier non-destructive testing (NDT) modality used for inspecting internal defects, particularly porosity in additive manufacturing and metal castings. The standard industry workflow relies on voxel-wise image segmentation (e.g., Otsu thresholding) to separate void spaces from the solid matrix before calculating the total Porosity Volume Fraction (PVF). However, this segmentation step is highly brittle under severe measurement noise and artifacts, leading to catastrophic miscalculations. In this paper, we propose a direct deep-learning regression approach using a 3D Convolutional Neural Network (CNN) equipped with Global Average Pooling (GAP) to predict scalar PVF directly from noisy 3D CT volumes, bypassing segmentation entirely. To validate this hypothesis, we engineered a rigorous synthetic 3D CT dataset containing 500 procedural phantoms with precisely controlled spherical pores and uniformly distributed PVF ranging from 0.1% to 15%. We degraded these volumes with varying levels of Gaussian noise (10 dB, 20 dB, and 30 dB SNR). The 3D CNN was evaluated against classical Otsu thresholding and a Random Forest (RF) machine learning baseline utilizing global intensity statistics. Our results demonstrate that the 3D CNN is exceptionally robust, maintaining a low Mean Absolute Error (MAE) of 0.363 percentage points even under extreme 10 dB noise, where Otsu thresholding catastrophically collapses (MAE of 11.51 pp). Unexpectedly, the RF baseline achieved the lowest overall MAE (0.088 pp). Through a rigorous evidence audit and domain-shift experiment involving a simulated beam-hardening cupping artifact, we dissect this phenomenon, revealing that the RF's superiority is a statistical artifact of the synthetic data's homogenous matrix assumption. This research highlights the viability of segmentation-free volumetric regression for NDT while providing a critical warning regarding 'shortcut learning' in synthetic physics datasets.
Authors
- Mayank Goswami (ORCID: https://orcid.org/0000-0001-8829-9980)
- Bhanu Prakash Godara
Institutions
- Indian Institute of Technology Roorkee (IN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-18
- DOI
- https://doi.org/10.21203/rs.3.rs-11733357/v1
- Primary Topic
- Advanced X-ray and CT Imaging
- Type
- preprint