Improving µCT image segmentation through architectural enhancements in the U-Net model
نویسندگان
1 Simulation and Data Processing Laboratory, School of Mining Engineering, College of Engineering, University of Tehran, Tehran, Iran.
2 Civil and Environmental Engineering Dept, School of Mining & Petroleum Engineering, Faculty of Engineering, University of Alberta, Edmonton, Alberta, Canada.
3 School of Mining Engineering, College of Engineering, University of Tehran, Tehran, Iran.
doi
10.22059/ijmge.2025.399223.595283چکیده
Micro-Computed Tomography (µCT) is an indispensable non-destructive technique for characterizing the internal microstructures of materials like porous media, but any quantitative analysis hinges on accurate image segmentation—a task complicated by image noise, poor contrast, and intricate features. While deep learning, and the U-Net architecture in particular, has shown considerable promise in automating this process, this study presents a comparative analysis between the standard U-Net and a bespoke, modified architecture for segmenting multi-phase micro-computed tomography (µCT) images of Bentheimer sandstone. Our modified U-Net introduces several architectural enhancements: optimized convolutional blocks for superior feature extraction, spatial dropout for more effective regularization, and L2 weight regularization to mitigate overfitting. We trained both models on 2D slices from two core samples and subsequently evaluated them against an independent blind test set of 1872 slices from a third core, which contained four distinct phases: porosity, quartz, clay, and feldspar. The quantitative results reveal that the modified U-Net decidedly outperforms its standard counterpart, achieving a macro-averaged Dice Similarity Coefficient (DSC) of 0.92 versus 0.88, and a macro-averaged Intersection over Union (IoU) of 0.85 versus 0.80. Most notably, our model demonstrated substantial gains in segmenting the challenging minority phases; the DSC for clay surged from 0.71 to 0.85, and for feldspar, it rose from 0.85 to 0.87, all while maintaining stable performance on the majority phases of porosity and quartz. These statistical improvements are corroborated by qualitative visual assessments, which confirm superior boundary delineation and a reduction in misclassifications. Ultimately, our findings indicate that the proposed architectural refinements yield a more accurate and robust segmentation model for micro-computed tomography (µCT) imagery, providing a more reliable foundation for downstream Digital Rock Physics (DRP) applications critical to the mining and geo-engineering sectors, such as geomechanical stability assessment and mineral liberation analysis.
