University of Twente Student Theses

Login

Generating high frame rate MRI images using a surrogate signal A Supervised Learning Approach

Shokry, Kirelloss (2018) Generating high frame rate MRI images using a surrogate signal A Supervised Learning Approach.

[img] PDF
1MB
Abstract:MR images are needed in many percutaneous minimally invasive procedures in the liver. However, the acquisition of MRI happens at low frequency, which doesn’t enable real-time imaging under respiratory induced motion in the liver. One main contribution of this study is that it presents a new technique for generating synthetic MR images that have contrast and spatial resolution higher than the acquired ones. The new approach depends on respiratory motion estimation in the liver using a surrogate signal. Synthetic MR images are predicted at the same high frequency of the surrogate signal. Also, this study presents a new machine learning algorithm for predicting the x-y pixel coordinates of small targets with very high accuracy and frequency without generating full images. The strength of this algorithm is that it tries to find the exact non-linear function between the surrogate signal and the position of targets in the MR image if it exists. Otherwise, the algorithm tries to capture the correlation between them as close as possible. The correctness of the first technique was evaluated by comparing the locations of liver features and the contours of the liver in synthetic images to their corresponding acquired ones. The mean error ranged between 0.659 pixel (1.3839 mm) and 1.2154 pixels (2.55 mm). The average blurriness of synthetic images was 0.3305, while that of corresponding acquired ones was 0.3306. The average entropy of synthetic images was 6.5631, while that of corresponding acquired ones was 6.5248. The mean error of the second technique was evaluated by comparing the predicted positions of targets to their real positions. The mean error ranged between 0.9859 pixel and 1.0034 pixels after training the algorithm by 50% of the acquired data. The mean error decreased to range between 0.8669 pixel and 0.8758 pixel after training the algorithm by 75% of the acquired images. The algorithm was tested every time by the last 25% of the obtained data. Both techniques were evaluated by open source data, previously carried out experiments and experiments carried out on a phantom.
Item Type:Essay (Master)
Clients:
Unknown organization, Enschede, Netherlands
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:01 general works
Programme:Embedded Systems MSc (60331)
Link to this item:http://purl.utwente.nl/essays/76692
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page