New methods for 3D image segmentation using Artificial Intelligence and Machine/Deep Learning.

Session

Poster Session Three

Authors

Jan Giesebrecht (1), Federico Gamba (1), Gwenole Tallec (1)

Affiliations

1. Thermo Fisher Scientific

Keywords

AI, Machine Learning, Deep learning, XCT, FIB-SEM, 3D, segmentation, imaging

Abstract text

Accurate 3D image segmentation is a fundamental step in image processing and analysis of a wide variety of biological, geological, and human-made materials. Artificial Intelligence (AI), a generic term for systems that interpret the environment and make accurate image processing decisions based on a set of ground truths, is becoming critical; it gives the advantage of being automated or semi-automated, and, perhaps most importantly, repeatable segmentation can be accomplished of multiple samples under the same conditions. Additionally, time if often a limiting factor, and AI can help with this. Here, we explore new available tools in Avizo Software: Machine Learning and Deep Learning as two domains of AI that can be used to achieve the required segmentation and subsequent analyses. 

Two examples are provided in this study: a focussed ion beam scanning electron microscopy (FIB-SEM) volume of shale, and an X-ray Computed Tomography (XCT) volume of a carbonate rock core. These are chosen to highlight the AI capabilities across different scales and modes. On the shale volume we apply a Machine Learning model to observe the 3D structure of pores and rock phases, achieved through supervised texture segmentation.  On the XCT carbonate core, we are interested in the 3D reconstructed pore space to understand the flow properties through the sample, and apply the same segmentation method. The challenges will be (a) to expand the AI approach across two very different samples, (b) to understand microporosity at different scales, (c) to identify the different minerals in each of the samples that could look like microporosity-impacted regions in different imaging modalities, and (d) the presence of fluid inclusions, which can look like other features at different scales. We will finish by discussing result validation and processing time, and the best method will be highlighted, considering the different constraints. Performance of various techniques will be quantitatively determined by using standard segmentation metrics as compared to a label volume that has been segmented manually. While understanding the benefits of AI through Machine and Deep Learning of geological samples is crucial for future exploration and energy requirements, this approach can also be applied a wide variety of biological and human-made materials. 

References