硅酸盐学报2026，Vol.54Issue(8)：2614-2626,13.DOI:10.14062/j.issn.0454-5648.20260295
基于半监督深度学习的混凝土骨料X射线计算机断层扫描图像分割与三维重构
Semi-Supervised Deep Learning for Aggregate Segmentation and 3D Reconstruction of Concrete from CT Images
摘要
Abstract
Introduction X-ray computed tomography(CT)is an important non-destructive technique for analyzing the microstructure of cement-based materials,which can characterize aggregate distribution,pore structure and interfacial transition zone.The accuracy of subsequent 3D reconstruction and quantitative analysis is highly dependent on the quality of image segmentation.However,concrete CT images generally have some problems such as low contrast between aggregate and matrix,blurred boundary and strong noise,which seriously limit the effect of conventional segmentation methods such as threshold method. Deep learning performs well in complex image segmentation,but its training relies on large-scale and high-quality labeled data sets,while CT image labeling is difficult and time-consuming.In practice,there are often only a few labeled slices,leading to insufficient generalization ability and unstable performance of the model,especially in low contrast and fine structure areas.To solve the contradiction between model complexity and data availability,it is necessary to develop efficient learning strategies that can make full use of unlabeled data. This paper was to propose a semi-supervised deep learning framework for automatic segmentation and three-dimensional reconstruction of concrete CT images.This method could effectively utilize a small number of labeled slices and a large number of unlabeled slices based on U-Net segmentation model and self-training mechanism.Through pseudo-label generation and confidence filtering,the boundary recognition and structural consistency under low contrast conditions could be improved. Methods The proposed framework consisted of three stages,i.e.,initial supervised training,pseudo-label generation,and iterative semi-supervised optimization. Firstly,U-Net was used as a backbone model,and its encoder-decoder structure was combined with jump connection,thus extracting high-level semantic and fine spatial features at the same time.The initial segmentation ability was obtained with 42 manually labeled CT slices,combined with Dice loss and binary cross entropy loss for training. Secondly,the trained model was applied to 2000 unlabeled CT slices to generate pixel-level prediction and probability maps.Through the confidence threshold,the high-reliability region was retained as a pseudo-label,and the low-confidence region(usually near the fuzzy boundary)was excluded.After morphological operation,the isolated noise was removed,and the boundary continuity was improved. Finally,the filtered pseudo-labels were merged with the original labeled data to form an augmented training set.The model was iteratively retrained on this mixed set,with labeled data providing stable supervision,and pseudo-labeled data supplementing structural information.Also,the learning rate attenuation strategy was introduced to ensure the stability and convergence of training.The Dice coefficient and the intersection-union ratio(IoU)were used to evaluate the segmentation performance. Results and discussion The performance of manual annotation quality and initial supervision model is evaluated.The marking mask can be close to the aggregate boundary in different scales,indicating great marking consistency and reliability.The U-Net model trained with only 42 labeled slices can quickly learn the discriminant characteristics of aggregate and matrix,and obtain a better verification performance(i.e.,the optimum verification Dice=0.860 6).However,in low contrast and complex structure areas,the performance still fluctuates,indicating that the generalization ability is limited when the training data is extremely limited. The semi-supervised learning strategy significantly improves the performance of the model.The pseudo-label visualization shows that the high confidence region is highly consistent with the actual aggregate distribution,and the uncertainty mainly appears at the aggregate-matrix interface.With iterative training,the indicators further improve.The Dice,IoU and accuracy of the final model(S-40)reach 0.971 2,0.944 0 and 0.982 0,respectively,which are 11.7%,22.7%and 7.4%higher than the fully supervised baseline,respectively,and the loss is reduced by 82.9%.These results show that high-confidence pseudo-labels can effectively enhance feature learning without compromising training stability. Three-dimensional reconstruction demonstrates that the proposed semi-supervised framework effectively preserves the spatial continuity and geometric integrity of aggregate particles.Compared with the fully supervised baseline,the reconstructed aggregates exhibit markedly fewer adhesion artifacts,better-defined particle boundaries,and more complete representations of fine aggregates and complex interfaces.These improvements enable reliable quantitative characterization of aggregate morphology,connectivity,and spatial organization,demonstrating that the proposed segmentation framework provides high-fidelity input for three-dimensional reconstruction. Quantitative analysis reveals that aggregate volumes follow a log-normal distribution and exhibit a pronounced surface area-volume scaling relationship(i.e.,A ∝V2/3),reflecting the geometric self-similarity of the reconstructed particles.Large particles contribute most of the total aggregate volume,whereas smaller particles occupy the interstitial spaces to form a dense skeleton-filling structure.The reconstructed aggregates are uniformly distributed throughout the specimen without evident segregation,and the extracted morphological descriptors yield a mean sphericity of 0.757,indicating moderate particle roundness while preserving angular characteristics.Collectively,these findings validate the reliability of the proposed reconstruction framework and provide a robust basis for quantitative three-dimensional microstructural characterization and aggregate gradation analysis. Conclusions This paper proposed a semi-supervised deep learning framework for concrete CT image segmentation and 3D reconstruction.This method effectively overcame the difficulties caused by limited labeled data and low imaging contrast via combining pseudo-label learning with U-Net architecture.The results showed that the model converged stably,the segmentation accuracy improved significantly,and an effective balance between manual labeling and high-precision microstructure analysis could be achieved.The framework could provide an extensible solution for the intelligent characterization of cement-based materials and have application prospects in the field of digital engineering.关键词
半监督学习/X射线计算机断层扫描图像分割/U-Net/伪标签/三维重建/水泥基材料Key words
semi-supervised learning/X-ray computed tomography image segmentation/U-Net/pseudo-label/3D reconstruction/cement-based materials分类
信息技术与安全科学引用本文复制引用
吉朵朵,范定强,吕学森,李旺,陆建鑫..基于半监督深度学习的混凝土骨料X射线计算机断层扫描图像分割与三维重构[J].硅酸盐学报,2026,54(8):2614-2626,13.基金项目
国家自然科学基金(52178249) （52178249）
广东省自然科学基金青年提升项目(2024A1515030243) （2024A1515030243）
武汉市知识创新专项(2023010201010094) （2023010201010094）
硅酸盐科学与先进建材全国重点实验室开放基金(SYYJJ2025-9). （SYYJJ2025-9）
