Work

Semantic Segmentation of CT Data of Oak Wood for the Creation of Wood Anatomical Climate Proxy Archives

Year: 2025

Type: article

Abstract: In this study, an automated method was developed for the semantic segmentation of computed tomography ( CT ) data of oak wood linking the fields of wood anatomy and image analysis. The method reliably and rapidly segments whole vessels regardless of the presence of tylosis, which frequently is a major problem in wood anatomical studies. Analysis of the segmented vessels can contribute to the expansion of wood anatomical climate proxy archives. A machine learning approach was used for the automatic segmentation of the vessels, and a neural network was identified that performs the segmentation task with the smallest possible error. To achieve this, different U-Net models — widely used deep learning architectures — were trained and evaluated with manually masked and augmented CT training data. As a peak performance of this dataset, an exactness of 0.79 measured at the mean intersection over union (IoU) could be achieved. The U-Net could generalize new CT data of oak wood and show good results in predicting CT -like image acquisitions, such as a digital scan of a thin section. By using the U-Net with the CT data, it is possible to quantify vessel diameter at extraordinarily small spatial increments and additionally observe the vessels in 3D. The trained U-Net is also integrated into a user interface into which single or multiple CT images can be automatically read, predicted, post-processed and saved. The code and model file will be permanently archived on Zenodo and GitHub (Füting 2025) and is open for further development.

Source: International Journal of Wood Culture

DOI: 10.1163/27723194-bja10056