Unsupervised segmentation of micro-CT scans of polyurethane structures by combining hidden markov random fields and a U-Net
Extracting a digital representation of a material from images is a prerequisite for any quantitative structure-property analysis. Supervised convolutional neural networks (CNNs) now deliver state-of-the-art segmentation accuracy, but their performance depends on large, manually annotated training se...
Salvato in:
| Autori principali: | , , , , , , |
|---|---|
| Natura: | Article (Journal) |
| Lingua: | inglese |
| Pubblicazione: |
7 February 2026
|
| In: |
Materials today. Communications
Year: 2026, Volume: 51, Pages: 1-10 |
| ISSN: | 2352-4928 |
| DOI: | 10.1016/j.mtcomm.2026.114817 |
| Accesso online: | Verlag, lizenzpflichtig, Volltext: https://doi.org/10.1016/j.mtcomm.2026.114817 Verlag, lizenzpflichtig, Volltext: https://www.sciencedirect.com/science/article/pii/S2352492826002011 
|
| Note sull'autore: | Julian Grolig, Lars Griem, Michael Selzer, Hans-Ulrich Kauczor, Simon M.F. Triphan, Britta Nestler, Arnd Koeppe |
| Riassunto: | Extracting a digital representation of a material from images is a prerequisite for any quantitative structure-property analysis. Supervised convolutional neural networks (CNNs) now deliver state-of-the-art segmentation accuracy, but their performance depends on large, manually annotated training sets—an impractical requirement for most bulk micro-computed-tomography (μCT) studies. Classical unsupervised techniques such as Hidden-Markov Random Fields (HMRF) avoid the need for ground-truth labels, yet they are typically slow and yield lower-quality segmentations. Here, we introduce HMRF-UNet, a hybrid framework that embeds the probabilistic neighborhood model of HMRF directly into the U-Net’s loss function. The loss simultaneously (i) enforces spatial smoothness through higher-order neighborhood terms, (ii) respects class-wise intensity distributions, and (iii) benefits from data-driven feature learning. By combining HMRF’s label-free regularization with the fast inference of CNNs, the method delivers unsupervised segmentation at a speed comparable to that of supervised networks. We evaluate the approach on a μCT dataset of polyurethane (PU) foam. An ablation study quantifies the contribution of each neighborhood term, and the HMRF-UNet attains a Dice similarity coefficient of 0.957±0.017, while processing a 256×256 slice in 200 ms on a single GPU; performance that rivals supervised baselines. To further diminish the reliance on annotated data, we propose a two-stage pre-training strategy: the network is first optimized with the HMRF loss on unlabeled data and subsequently fine-tuned on a minimal labeled subset. This approach recovers 98.4% of the fully supervised performance while using only 1% of ground-truth annotations. The proposed framework provides accurate, high-throughput segmentation without extensive manual labeling, enabling rapid, data-driven characterization of complex porous architectures across a broad range of material systems. |
|---|---|
| Descrizione del documento: | Gesehen am 16.06.2026 |
| Descrizione fisica: | Online Resource |
| ISSN: | 2352-4928 |
| DOI: | 10.1016/j.mtcomm.2026.114817 |