Making Ultra-Low-Field (ULF) MRI a reality

There are two types of modalities (QH) used for (high-resolution) medical imaging: MRI and CT (Computed Tomography scan using X-rays). The main disadvantage for CT is that X-rays are a type of ionizing radiation which can damage DNA and cause mutations. Because of the sensitivity of the CT method, only a small dose is delivered that is essentially harmless, but one must be careful not to accumulate too many doses. MRI poses no such danger. The magnetic field and radiofrequency pulses are not ionizing radiation and do not harm the body. In terms of imaging capability, MRI is better able to identify certain tumors because of superior contrast to the surrounding normal tissue. But because of its speed, CT can be better for scanning the whole body for any cancer metastases.

Magnetic Resonance Imaging (MRI) operates on the principles of nuclear magnetic resonance (NMR). The process involves several key technical steps:

• Magnetic Field Application: The patient is placed within a strong, static magnetic field, typically ranging from 0.2 to 3 Tesla. This field aligns the magnetic moments of hydrogen nuclei (protons) in the body's water and fat molecules along the direction of the applied magnetic field.
• Radiofrequency (RF) Pulses: The MRI machine emits RF pulses at a frequency specific to the resonance frequency of the hydrogen nuclei (Larmor frequency). The RF pulses temporarily perturb the alignment of the protons, tipping them away from the static magnetic field direction.
• Resonance and Relaxation: When the RF pulse is turned off, the protons gradually relax back to their original alignment with the static magnetic field. During this relaxation process, they emit RF signals. Two types of relaxation occur: 
• T1 Relaxation (Longitudinal Relaxation): The time it takes for the protons to realign with the static magnetic field.
• T2 Relaxation (Transverse Relaxation): The time it takes for the protons to lose phase coherence among each other, causing the signal to decay.
• Signal Detection: The emitted RF signals are detected by receiver coils. Spatial encoding is achieved using gradient magnetic fields in different directions (x, y, and z) to encode spatial information.
• Image Reconstruction: The detected signals, which are in the frequency domain, are converted to spatial domain images using the Fourier transform. The resulting data provides detailed images of the internal structures of the body, based on the varying signal intensities that correspond to different tissue types and their properties.

In summary, a strong magnetic field align protons in the water of body tissues, which are then perturbed by RF pulses. There are two types of relaxation signals (T1 and T2) that are detected and used to reconstruct images of the body.

MRI takes advantage of how T1 and T2 relaxation times vary in different tissues and tissue environments (e.g. brain versus skull or cancer versus non-cancer). This gives rise to two basic imaging modalities:

• T1-Weighted Imaging: Emphasizes T1 relaxation properties and is best for anatomical detail; it can be enhanced using contrast dyes. Tissues with long T1 relaxation times, such as cerebrospinal fluid (CSF), appear dark because they take longer to realign and thus contribute less signal.
• T2-Weighted Imaging: Emphasizes T2 relaxation properties and is best for detecting pathologies; it is sensitive to water content, and can detect increased water content such as edema and inflammation.

A strong magnetic field is important because it results in a stronger MRI signal; the stronger the magnetic field, the greater the difference in energy levels between the aligned and anti-aligned protons, leading to a stronger signal that can be detected by the MRI receiver coils. Indeed the signal is proportional to the magnetic field strength squared.

Thus, a more powerful magnetic field results in an improved signal-to-noise ratio (SNR) which means clearer and more detailed images. High SNR is crucial for distinguishing between different types of tissues and identifying abnormalities, such as tumors, lesions, or other pathological conditions. In addition, improved SNR allows for higher spatial resolution. Higher resolution images enable the visualization of finer anatomical details, which is critical for accurate diagnosis and treatment planning.

Finally, stronger magnetic fields allow for faster imaging sequences. This means that scans can be completed more quickly, reducing the time the patient needs to remain still in the MRI machine, which can improve patient comfort and reduce motion artifacts in the images.

Magnetic Resonance Imaging (MRI) machines typically use magnetic fields ranging from 0.2 to 3 Tesla (T), with most clinical MRI scanners operating at 1.5 T or 3 T. Such high field strengths are achieved through superconducting magnets that require expensive liquid helium to keep cool. A permanent magnet made from ferromagnetic material can operate at room temperature but generate weaker fields. As a point of comparison, the bar magnets found in schools create field strengths of the order of 1 × 10^-4 Tesla.

Because MRI machines are expensive and need special facilities, access can be limited especially in poorer countries. This leads to unequal distribution and limited availability of MRI scans. As described in a new paper in Science, a research group in Hong Kong have constructed a highly simplified whole-body ultra-low-field (ULF) MRI scanner that uses a compact 0.05 Tesla permanent magnet to generate the magnetic field. 

One way to compensate for the poorer signal-to-noise ratio and lower resolution from a weaker magnet is employing a more sensitive detector with better noise filtering properties. Instead, the Hong Kong group opted for a cheaper software approach using deep learning reconstruction methods. They trained their deep learning network to take a low-resolution image as input and transform it into a high-resolution image as if it were obtained from a high-magnetic field MRI scan. Thus, the training and test sets consisted of lower resolution 0.05 Tesla images paired with 3 Tesla images from the same subject.

In this manner, starting with 0.05 T data acquired at 3 mm (low) resolution, they were able to generate 1 mm (high) resolution images improving image quality and restoring fine neuroanatomical structures (see Figure 1). In the paper, the authors state that the method enabled clear identification of structural details in organs such as liver, kidneys, stomach, pancreas, spleen, and spine. In the future, one may expect to see more examples of AI software systems helping to overcome hardware limitations. 

As with any machine learning algorithm, the ultimate test is when the ULF scanner is used widely in the real world. Often there is a performance decay as the system encounters unexpected or novel input data (e.g. noisier images in clinic setting) significantly different from the data the method was originally trained and tested on. These out-of-distribution (OOD) data may cause the algorithm to reconstruct an inaccurate high-resolution image. So it is important for the ULF MRI to have its accuracy periodically re-evaluated.

But the benefits of a cheaper more portable MRI scanner are substantial which the authors highlight in their conclusions:

"Moreover, we demonstrated the potential of deep learning image formation to substantially augment 0.05 Tesla image quality by exploiting computing and extensive high-field MRI data. These advances pave the way for affordable, patient-centric, and deep learning–powered ULF MRI scanners, addressing unmet clinical needs in diverse healthcare settings worldwide."

Figure 1. Before and after deep learning reconstruction  and enhancementof 0.05 T MRI images. For each image pair, the left image represents the raw output image from Ultra- MRI scanner, and the right image represents the higher-resolution image after deep learning image processing. Anatomical details are seen more clearly in the image on the right (from Zhao et al. Science, 2024).