Unsupervised Segmentation of Micro-CT Scans of Polyurethane Structures By Combining Hidden-Markov-Random Fields and a U-Net

TL;DR

This work presents a method that integrates HMRF theory and CNN segmentation, leveraging the advantages of both areas: unsupervised learning and fast segmentation times, and proposes and demonstrates a pre-training strategy that considerably reduces the required amount of ground-truth data when training a segmentation model.

Abstract

Extracting digital material representations from images is a necessary prerequisite for a quantitative analysis of material properties. Different segmentation approaches have been extensively studied in the past to achieve this task, but were often lacking accuracy or speed. With the advent of machine learning, supervised convolutional neural networks (CNNs) have achieved state-of-the-art performance for different segmentation tasks. However, these models are often trained in a supervised manner, which requires large labeled datasets. Unsupervised approaches do not require ground-truth data for learning, but suffer from long segmentation times and often worse segmentation accuracy. Hidden Markov Random Fields (HMRF) are an unsupervised segmentation approach that incorporates concepts of neighborhood and class distributions. We present a method that integrates HMRF theory and CNN segmentation, leveraging the advantages of both areas: unsupervised learning and fast segmentation times. We investigate the contribution of different neighborhood terms and components for the unsupervised HMRF loss. We demonstrate that the HMRF-UNet enables high segmentation accuracy without ground truth on a Micro-Computed Tomography ($μ$CT) image dataset of Polyurethane (PU) foam structures. Finally, we propose and demonstrate a pre-training strategy that considerably reduces the required amount of ground-truth data when training a segmentation model.

Figures (7)

• Figure 1: Overview of the Segmentation U-Net Architecture
• Figure 2: Visualization of the proposed cuboid-based dataset split. By extracting non-overlapping cuboids and splitting the data based on the cuboids, a leakage between training, validation and test set can be prevented.
• Figure 3: Comparison of segmentation results for different images from the testing set for five models trained with different neighborhood terms. The models trained with neighborhood term (columns 2-6) were trained with a neighborhood weight of $\lambda_n=0.31$. The model without neighborhood term (second column) over-segments the PU structure inside the pore space (blue areas). Training with the neighborhood term removes these over-segmentations. For the normal Banerjee neighborhood term an over-segmentation around the PU walls can be observed for the worst test image in the bottom row (blue areas). Otherwise most segmentation errors resulted from under-segmented PU structures (read areas).
• Figure 4: Comparison of segmentation results of models trained with weighted Banerjee neighborhood with different thresholds $\sigma_{\text{thresh}}$ for two exemplary images. In the regions marked with red circles, differences of segmentation accuracy become apparent, showing improved segmentation of thin structures for smaller thresholds.
• Figure 5: Effect of using HMRF-UNet as a pretrained model for finetuning with different amount of ground-truth pairs using supervised loss. The grayscale image (first column), the ground-truth segmentation (second column), the segmentation of the supervised model without pretraining (third column) and the segmentation of the supervised model with HMRF-UNet pretraining are shown for examples of the test set with the smallest change (first row), median change (second row) and biggest change (third row). The colour coding of the segmentations in the third and fourth column show the improvement of using more images for fine-tuning.