Automatic Object Detection in Images

I have a TIFF stack representing 3-D grayscale imaging data of biomass char particles with support material. There are a total of 2016 2-D images, where each image slice is 2006 pixels by 2006 pixels. When you look at the full 3-D volumetric data, you will see elliptical particles (I want to keep them) and spherical support material (I want to remove them). I have attached a sample slice from the middle of the volumetric stack (i.e., Slice 1008 in the x-y plane, namely the top view).

https://postimg.cc/hXJCwgwd

I have tried numerous morphological operations (e.g., Otsu's Method, Opening, and Recursive Gaussian Filtering) and other tools (e.g., Hough Transform) to remove the undesired support material, but I notice that the tools remove regions from the desired elliptical particles. I have tried to apply both Machine Learning (e.g, Support Vector Machines) and Deep Learning (e.g., Convolutional Neural Networks) techniques by training and testing on the same image, which should correctly detect the undesired support material and allow me to subsequently remove it as a post-processing step. Unfortunately, I have not been able to remove only and all of the undesired support material and save the desired elliptical particles.

As you can see from the sample slice, there are a lot of issues with detecting and eventually removing the undesired spherical support material:

* Particles and support material are not exactly elliptical (particles) and spherical (support material) in shape.

* No well-defined edges on both the elliptical particles and spherical support material.

* Elliptical particles and spherical support material are not solid inside their structures.

* No overlap but touching between the elliptical particles and spherical support material.

* Elliptical particles and spherical support material appear and disappear as you move through all 2016 2-D slices of the 3-D volumetric data.

What are my options on automatically detecting and removing the spherical support material but keeping the elliptical particles?

I appreciate any feedback from the community.

Answer 1

[Ronit]

Hello,

I would recommend you try out the following MATLAB tools and suggestios to achieve your goal:

1. For 3D Image Segmentation, you can leverage the Image Processing Toolbox. It provides functions for reading TIFF stacks ('imread', 'tiffreadVolume'), filtering ('imgaussfilt3'), and thresholding ('imbinarize', 'graythresh').
2. Use 'bwconncomp' and 'regionprops3' to label and analyze individual objects in 3D.
3. After segmentation, for shape analysis for each 3D object's, you can leverage the 'regionprops3' fuinction. It can compute properties such as 'PrincipalAxisLength', 'Volume', and 'SurfaceArea'. Remove objects based on their shape metrics.
4. You can leveraage the Deep Learning Toolbox for semantic segmentation. Please refer to the following link for more details regarding this: https://www.mathworks.com/help/vision/ug/semantic-segmentation-using-deep-learning.html

Refer to the following documentation pages for further details regarding the above descriobed functions and toolboxes:

• Image Processing Toolbox: https://www.mathworks.com/help/images/index.html
• Deep Learning Toolbox: https://www.mathworks.com/help/deeplearning/index.html
• 'imread': https://www.mathworks.com/help/matlab/ref/imread.html
• 'tiffreadVolume': https://www.mathworks.com/help/images/ref/tiffreadvolume.html
• 'imgaussfilt3': https://www.mathworks.com/help/images/ref/imgaussfilt3.html
• 'imbinarize': https://www.mathworks.com/help/images/ref/imbinarize.html
• 'graythresh': https://www.mathworks.com/help/images/ref/graythresh.html
• 'bwconncomp': https://www.mathworks.com/help/images/ref/bwconncomp.html
• 'regionprops3': https://www.mathworks.com/help/images/ref/regionprops3.html

I hope the above suggestions help you resolve the query!