LERSNet: A lightweight CNN-Transformer hybrid for high-accuracy multi-mineral segmentation in micro-CT rock images

This is the authors' abstract. We don't add key points for this paper.

Accurate segmentation of rock constituents in micro-CT images is crucial for predicting petrophysical properties. While deep learning-based methods have improved segmentation performance, they often struggle with geologically complex materials, where subtle attenuation differences and textural heterogeneity challenge conventional architectures. In addition, existing architectures often fail to strike a balance between accuracy and computational efficiency, particularly for complex mineral assemblies in clay-rich sandstones. This study introduces LERSNet (Lightweight Efficient Rock Segment Network), a purpose-built lightweight CNN-Transformer hybrid network optimized for multi-mineral segmentation. It uniquely integrates strong, complementary modules including positional encoding attention, multi-scale attention, depthwise separable convolutions, channel attention, and residual asymmetric convolutions, along with transformer-based global context modeling, to resolve overlapping mineral attenuations. We benchmark LERSNet against seven state-of-the-art CNNs and traditional Otsu thresholding using a meticulously labeled micro-CT dataset of Bentheimer sandstone (pore, quartz, clay, feldspar). LERSNet achieves a composite F1-score of 96.1% (vs. 80.0–95.2% for other CNNs) and 87.6% on blind-test volumes, with superior clay segmentation (86.9% F1-score) due to its edge-guided fusion and positional encoding. Despite its high accuracy, LERSNet maintains computational efficiency (8.33 GFLOPs, 0.7M parameters), enabling practical deployment. Our results demonstrate that hybrid local-global feature learning, combined with geologically informed training protocols, significantly outperforms conventional methods, advancing automated rock microstructure analysis for petrophysical applications.