Original Source Here
Low Resolution MRI for Deep Learning Models
A lot of MRI datasets are available for training deep learning models related to different task such as segmentation of body organs, detecting abnormalities and many more use cases. This blog mainly focuses on the models which try to enhance MRI scans(Low resolution). The main objective of such models is to reduce the scan time. The process will be to get a fast MRI Low Resolution(LR) scan and the pass it through a deep learning model which will give a Super Resolution(SR) thus reducing overall scan time.
For training a deep learning model for generating SR from LR MRI scans we need a dataset which will have pairs of High Resolution(HR) MRI Scans and Low Resolution MRI Scans for same object generated using same scanner. There are lot of open source HR MRI datasets available for different body parts, but we don’t have the corresponding LR scans for same dataset. The only option we have is to generate the LR MRI scans using available HR scans.
Now before diving into the process of generating Low Resolution images lets get an overview of what is actually a MRI scan and some concepts related to it.
Magnetic Resonance Imaging (MRI) is a non-invasive imaging modality to have an insight into the human body. In MRI the object to be imaged is placed in a strong magnetic field, known as Bo field, and a Radio Frequency (RF) signal, called B1 field, is applied to selectively excite the parts of the object to be imaged. The signals from the excited slices of the object are obtained using receiver coils. The data received from these receiver coils is stored in k-space (raw data space of MRI). The inverse Fourier transform of the k-space provides the Magnetic Resonance (MR) image.
The key takeaway form above description about MRI Scan is raw MRI data is stored in the form of something called as k-space.
k-space is an abstract concept and refers to a data matrix containing the raw MRI data. The k-space is an extension of the concept of Fourier space well known in MR imaging. It represents the spatial frequency information in two or three dimensions of an object. The k-space is defined by the space covered by the phase and frequency encoding data. The relationship between k-space data and image data is the Fourier transformation. Below are some of the important properties of k-space data –
- The size of the k-space is the same as the size of the MR image
- A point in the k-space does not correspond to a point in the image matrix.
- The data at the center of the k-space contains most of the signal information and the contrast information of the image, and the data along the outer edges contains information about the edges and boundaries.
Now we have just enough understanding about MRI scan and k-space. Let us begin with our main task of generating Low Resolution Scans using High Resolution MRI scans.
Generating Low Resolution MRI
The low resolution images can be generated in image domain or in k-space. There are plenty of traditional image processing techniques available which can be used to generate LR image. Here we will focus mainly on k-space sub-sampling techniques. Most of the techniques around filtering the k-space data. This techniques can be applied to both 2D and 3D data. If we have raw k-space data from scanner we can manipulate the k-space data directly to generate LR k-space and the we can apply inverse Fourier transform (iFFT) to get an MR image. If we have a MR image then we need to convert it to k-space then apply below transformations and the again apply iFFT to get LR MRI.
Below is a sample HR brain image for comparison-
Left side image is image domain visualization and right side is k-space image.
Trending AI/ML Article Identified & Digested via Granola by Ramsey Elbasheer; a Machine-Driven RSS Bot