# AI-assisted deep learning segmentation and quantitative analysis of X-ray microtomography data from biomass ashes

## TL;DR

This paper introduces a deep learning method to analyze the 3D structure of biomass ash particles using X-ray microtomography, improving accuracy and efficiency over manual methods.

## Contribution

A U-Net-based deep learning model is proposed for accurate segmentation and quantitative analysis of ash particle microstructures.

## Key findings

- The deep learning model successfully segmented ash and pores in heterogeneous particles.
- Quantitative metrics like porosity, pore size distribution, and sphericity were accurately extracted.
- The model resolved segmentation challenges caused by similar intensities and background artefacts.

## Abstract

X-ray microtomography is a non-destructive method that allows for detailed three-dimensional visualisation of the internal microstructure of materials. In the context of using phosphorus-rich residual streams in combustion for further ash recycling, physical properties of ash particles can play a crucial role in ensuring effective nutrient return and sustainable practices. In previous work, parameters such as surface area, porosity, and pore size distribution, were determined for ash particles. However, the image analysis involved binary segmentation followed by time-consuming manual corrections. The current work presents a method to implement deep learning segmentation and an approach for quantitative analysis of morphology, porosity, and internal microstructure. Deep learning segmentation was applied to microtomography data. The model, with U-Net architecture, was trained using manual input and algorithm prediction.•The trained and validated deep learning model could accurately segment material (ash) and air (pores and background) for these heterogeneous particles.•Quantitative analysis was performed for the segmented data on porosity, open pore volume, pore size distribution, sphericity, particle wall thickness and specific surface area.•Material features with similar intensities but different patterns, intensity variations in the background and artefacts could not be separated by manual segmentation – this challenge was resolved using the deep learning approach.

The trained and validated deep learning model could accurately segment material (ash) and air (pores and background) for these heterogeneous particles.

Quantitative analysis was performed for the segmented data on porosity, open pore volume, pore size distribution, sphericity, particle wall thickness and specific surface area.

Material features with similar intensities but different patterns, intensity variations in the background and artefacts could not be separated by manual segmentation – this challenge was resolved using the deep learning approach.

Image, graphical abstract