Deep Learning-Based Segmentation of 3D Volumetric Image and Microstructural Analysis.

Area of Science:

• Computer Vision
• Deep Learning
• Materials Science

Background:

• 3D object segmentation is crucial for applications like medical imaging and autonomous vehicles.
• Traditional methods struggle with accuracy and generalization for complex 3D data.
• Deep learning, particularly Convolutional Neural Networks (CNNs), has shown superior performance in 2D computer vision tasks.

Purpose of the Study:

• To develop an accurate and efficient method for multiclass 3D segmentation of volumetric data.
• To analyze the microstructures of sandstone samples by segmenting and characterizing individual components.
• To demonstrate the effectiveness of a combined 3D UNET and VGG19 architecture for detailed particle analysis.

Main Methods:

• A CNN-based architecture, 3D UNET, inspired by 2D UNET, was employed.
• The model was enhanced by integrating VGG19 for multiclass segmentation of sandstone microstructures.
• 448 2D images were aggregated into a 3D volume for analysis, with subsequent particle characterization using IMAGEJ.

Main Results:

• The proposed method achieved a high accuracy of 96.78% and an Intersection over Union (IOU) of 91.12% in segmenting sandstone microstructures.
• The model successfully segmented four distinct object classes within the volumetric data.
• Detailed analysis of individual particles, including average size and area percentage, was performed.

Conclusions:

• Deep learning models, specifically the 3D UNET and VGG19 combination, are highly effective for analyzing volumetric data and sandstone microstructures.
• The developed method offers computational insights for real-time implementation and surpasses current state-of-the-art techniques.
• This approach provides a robust model for microstructural analysis of various volumetric datasets.