Talk Details

Presenter:

Title:

Authors:

Abstract:

As micro-CT systems are becoming more prevalent and accessible, it is increasingly possible to  collect large quantities of 3D rock data. However, the time taken to analyse those data is still a significant bottleneck for researchers, often requiring postgraduates and post-doctoral researchers to invest significant time into image segmentation. Whilst semi-automated methods, such as watershed and top-hat segmentation, have helped to reduce the time burden of analysing 3D data, they aren’t suitable for every sample and still require manual adjustments to achieve high-quality segmentations for each sample. A new solution to this problem is the development of AI deep learning segmentation such as U-Net and machine learning such as Random Forest. This technique has many potential benefits over manual and semi-automated techniques, such as leveraging spatial context, texture, and geometric features rather than relying solely on grayscale intensity. Once trained, these models offer the potential for near-instantaneous segmentation of large, similar datasets. However, the limitations of AI, specifically the time required for generating high-quality training data and the potential for model hallucination, remain valid concerns for users. To quantify the efficacy of these tools, this study compares manual, semi-automated, and AI-driven workflows using real-world geological samples. The performance of each method is benchmarked by measuring the total time investment against segmentation accuracy, quantified using intersection over union and dice similarity coefficients. I will also delve into the future of AI-driven segmentation using pre-trained models, the limitations of each method and how to decide which machine learning or deep learning architecture work best for different data.

Alert

page: /