New Deep Learning Method enhances MRI outcomes

Deep Learning is a cluster of machine learning techniques and is implemented through a neural network. Therefore, it is also known as a deep neural network. Unlike Machine Learning , it uses multiple hidden layers to continue extracting high-level features from the raw input. It reflects the intricacies of the human brain, retrieves information and performs decision making.

 

From Natural Language Processing (NLP), fraud detection, entertainment (Netflix, sports etc.), Deep Learning has surpassed its horizon and recited an incantation to explore the medical field.

 

(Related Reading: Role of AI in Medical Field)

 

Research

 

A team of researchers has recently derived a new deep learning method from their research on the minimization of artifacts (the blurred spots that are created if image quality is compromised due to a movement) and noise that usually add themselves due to the slightest movement and also for a brief image-acquisition time. 

 

A professor of radiology at the School of MIR, Hongyu An and Ulugbek Kamilov led a research team that generated a deep learning method, which was trained by using images with artifacts.

 

(Recommended Blog: Deep Learning Algorithms)

 

Kamilov explained the MRI process, saying that it depends on someone's health if it turns out to be easy or difficult for that person to be scanned. 

 

But one has to continue breathing, and while someone does so, the internal organs make a movement and this is what, as he adds, "we have to determine how to correct for those movements.”

 

A doctoral student in Kamilov's lab, Weijie Gan penned down the software for Phase2Phase so that it eradicates noise as well as artifacts. An instructor of radiology at MIR, Cihat Eldeniz researched on the MRI acquisition-time and motion detection that are put into effect in the study.

 

According to TodayUKNews, during the process of Phase2Phase, the research team feeds the deep learning model with a cluster of bad images and trains it accordingly to provide a stark contrast between a good image and a bad image without providing a ground truth reference.

 

Retrospectively, the team assessed MRI data from 33 people, among which 15 were healthy and the rest were down with liver cancer. All these patients were allowed to breathe normally. The outcome was compared with images retrieved with UNet3DPhase.The Phase2Phase method has assessed 66 MRI datasets.

 

However, two radiologists were blinded in order to determine which reconstruction method was used during the process. It was observed that The Phase2Phase and UNet3DPhase images were quite similar with respect to their sharpness and contrast. And the UNet3DPhase images had fewer artifacts as compared to Phase2Phase.'

 

It was also observed that the Phase2Phase images keep the motion vector fields intact, unlike compressing sensing images.

 

Conclusively, An stated that the Phase2Phase deep learning method provides a "rapid reconstruction of high-quality 4D liver images using only a fraction of acquisition time."