SAM-IQ: Instance segmentation for 3D image data of particulate systems

Journal article, 2026

Accurate 3D instance segmentation is essential for quantitative analysis in engineering and science, especially in systems like clay suspensions where particles are sub-micron, anisotropic, and densely packed. In this study, we introduce SAM-IQ: a spatial 3D implementation of Segment Anything Model 2 (SAM2). We then compare it against two other segmentation strategies applied to a synchrotron X-ray nano-CT image of Speswhite kaolin clay in suspension: (i) a traditional image processing pipeline (Filtering Distance map-based Segmentation, FDS) and (iii) SAM-IQ-clAI, a fine-tuned variant of the model trained on a clay-specific dataset of annotated 2D slices (clAI-D). Performance was assessed using voxel-level Dice similarity and particle morphology descriptors such as size, flatness, and elongation. Results reveal that each method yields significantly different particle counts and shape metrics, highlighting that segmentation choices directly influence particle statistics and modelling outcomes and underscoring the need to treat segmentation as an integral part of measurement rather than a neutral preprocessing step. SAM-based methods consistently outperform FDS in dense, low-contrast volumes, with fine-tuning improving detection of sub-micron particles but yielding only marginal gains in morphological accuracy. By combining automated prompt generation, adaptive termination, and minimal training requirements, SAM-IQ provides a scalable, data-efficient framework for 3D segmentation in noisy volumetric images, with applications extending well beyond clay systems.