Epipolar-Constrained Optical Flow Triangulation for the Interior Problem in CBCT
To access to the full paper please contact me.
Epipolar-Constrained Optical Flow Triangulation for the Interior Problem in CBCT
In 2019 I worked as a research assistant at OVGU and helped with the above-mentioned research project. This research was published in 2020 and I was the co-author of this research paper. In this section, I introduce the topic and discuss my responsibilities in this research.
Introduction
The reconstruction of the truncated CT projections is a challenging task. The reconstruction of the projections can only be done for the body parts which were available in the CT field of view. If the field of view is completely embedded within the patient this is called the interior problem. In such cases, the data is truncated, and discovering the shape and the size of the object being scanned is very difficult. In this research, a method is represented which allows obtaining knowledge about the size and shape of the scanned object. In this method, optical flow and the epipolar geometry are combined to find the correspondences between different frames of the acquisition (e.g., the bony structure of the nose) and triangulate them back to their projective points of origin.
Objectives
Using this method, the objective is to see how much shape information can be obtained by considering rotational movement on acquisition trajectory between consecutive projection frames. Considering truncated data, this information can be used to acquire spatial information outside of the field of view which can also be used to estimate the size of the patient’s head.
The main goal of my scientific work in this research was to optimize the codes (written in MATLAB) in a way that the markers in the reconstructed image efficiently localize the point of interest that were chosen in the projection data. This included pre-processing of the CT data and optical flow method, and optimization of the CT image's window center and width values.
Methods
The main focus of my task was to minimize the noise, reduce the number of markers misplacement and efficiently increase the number of markers that specifically localize bony structures. To obtain the goal, the following changes were applied to the code:
1- I optimized the window center and window width using FIJI. For each slice of a single data, the optimized window level and width were acquired by FIJI window/level adjustment. The optimized values were then averaged over 11 slices of a single data to acquire an optimized average. The same procedure was done for all the other data and finally, an optimized average has been acquired for 6 datasets.
2- In the second stage different filters and methods were used to efficiently remove the noise while at the same time preserving the sharpness of the images. The following filters were tested in this study:
Results
The number of obtained localized markers depends on several factors including the type of filter, window center and window width, appropriate, thresholding, window size, etc. An appropriate trade-off between these parameters is required to obtain an efficient result. After applying the filtering methods discussed in the previous section to the data, the following results were acquired:
• Bilateral filtering which is an edge-preserving filter and is widely used in CT data for denoising and edge-preserving purposes.
• Image-guided filtering which is also an edge-preserving filter
• Direct high pass filter using pad array in MATLAB
• Median filter
• Smoothing the image using the average filter for denoising purposes which was followed by the application of Imsharpen in MATLAB to sharpen the data
• Smoothing the image using a gaussian filter followed by application of Imsharpen in MATLAB
• Sharpening the image using Imsharpen followed by the application of image-guided filter
• Sharpening the image using Imsharpen followed by the application of the median filter
•Sharpening the image using Imsharpen followed by the application of bilateral filter
3- In the third stage based on the acquired result in stage 2, optimization of the optical flow thresholding and window size of the local max function was performed to obtain the best-localized markers.
• Edge preserving filters decreased the noise but did not efficiently detect the edges
• Using the Imsharpen function alone or applying a direct high pass filter without the application of smoothing, increased the noise and the misplacement of the markers. Furthermore, edges and bony structures were not efficiently captured
• To decrease the noise and increase the number of markers that efficiently detect the bony structures, another approach was used. The pre-processed projection was initially smoothened using the filters that had better edge-preserving results. The filtered image was then sharpened using the Imsharpen function in MATLAB.
• The best marker localization was obtained using image-guided filtering method followed by the application of Imsharpen. Chick and nose bones were localized efficiently with markers.
• Acquired result should be applicable to the whole dataset, therefore adjustment of the window center, width, window size in local max, and appropriate thresholding of optical flow was necessary. The best values were acquired experimentally.
During this project, I obtained vast knowledge about image processing with MATLAB and learned how to write a scientific report.