A Deep Learning Approach Using Masked Image Modeling for Reconstruction of Undersampled K-spaces

Increasing the efficiency of a magnetic resonance imaging (MRI) scanner can save time, money, and most importantly lives (Jahnke 2019). Most modern day MRIs are effective but have the same major drawbacks; they are expensive and time consuming (Ghadimi et al 2022). The cost can largely affect overall access to MRIs, while the time consumption has a multitude of major effects. Additionally, for patients with pre-existing conditions such as claustrophobia, the time spent in the machine could have negative mental and physical effects, as well as reducing image quality due to the likelihood the patient would move during screening. Patients with certain contraindications such as pacemakers, artificial limbs or hearing aids can only enter MRI machines for a short period of time, meaning the accuracy of their images would be significantly lower than necessary (Ghadimi et al 2022). With the technology of the given age, MRIs are the next target for improvement. Recent advances in machine learning (ML) have sparked the development of artificial intelligence (AI) in medicine. The expanding field of disease detection gave way to many studies focusing on observable characteristics in imaging outputs like x-rays or MRIs. For example, a study in 2020 used a VGG-16 transfer learning model to evaluate its performance for detecting COVID induced pneumonia (M. D. Hasan et al 2021). Then came the development of the modern transformer architecture, such as Google’s BERT model. These models were revolutionary in the sense that they learned through context rather than patterns like previous models (Khan et al 2022). Though these models were revolutionary, their advancement also created an innovation break in a rising technique called masked image modeling (MIM). This occurred because most