Speaker
Description
Three-dimensional multiphase segmentation of pore-scale X-ray CT imagery in porous media faces a fundamental bottleneck that extends beyond achieving high in-domain accuracy on individual volumes. A key limitation lies in the absence of artificial intelligence methods that can function as unified segmentation models across multiple samples. Existing deep learning approaches for porous media segmentation often suffer from pronounced domain shift when variations arise in rock type, imaging system and acquisition parameters, or fluid-bearing conditions. Consequently, models typically require retraining or repeated fine-tuning for each new sample, which substantially increases both annotation effort and computational cost. This sample-specific training paradigm restricts the scalability and reusability of AI-based segmentation within digital rock analysis and pore-scale multiphase flow imaging workflows.
To address these challenges, we propose Mamba‑UNet, an efficient 3D segmentation framework built around State Space Models (SSMs), designed to improve cross-sample and cross-scanner generalization while maintaining computational efficiency. We develop a micro‑CT–specific augmentation strategy to better account for intrinsic noise and structural variability, and to emulate shifts in imaging conditions and intensity statistics. We further introduce a tri-orientated scan collaboration module to capture long-range spatial dependencies and global contextual information throughout the volumetric domain. In addition, an uncertainty estimation mechanism is incorporated to adaptively assess feature reliability during multi-scale fusion, enhancing fusion robustness under domain shift.
The proposed Mamba-UNet framework is evaluated on publicly available Bentheimer sandstone and Ketton carbonate datasets. Experimental results demonstrate that the model achieves competitive segmentation performance and efficient inference on these benchmarks, while also maintaining strong segmentation quality on an unseen Bentheimer sandstone dataset excluded from training. Furthermore, the method exhibits stable performance on fluid-bearing Bentheimer sandstone and Ketton carbonate volumes acquired using different imaging systems. These results highlight the reusability and scalability of the proposed approach for multi-sample digital rock workflows, providing more reliable 3D segmentation to support high-throughput pore-structure quantification and pore-scale multiphase flow studies.
| Country | United Kingdom |
|---|---|
| Student Awards | I would like to submit this presentation into both awards |
| Acceptance of the Terms & Conditions | Click here to agree |
