Portable MRI and AI improve Alzheimer’s diagnosis with cost-effective precision

Revolutionizing dementia care: Discover how portable, AI-driven MRI systems are breaking barriers in Alzheimer’s diagnosis, enabling early detection and global accessibility.

Study: Portable, low-field magnetic resonance imaging for evaluation of Alzheimer’s disease. Image Credit: illustrissima / Shutterstock

A recent Nature Communications study optimized the portable LF-MRI acquisition and developed a machine learning pipeline to estimate brain morphometry and white matter hyperintensities (WMH) for Alzheimer’s disease diagnosis.

Alzheimer’s disease (AD): Pathology and diagnosis

AD is a progressive neurodegenerative disease that affects memory, thinking, and behavior. It is pathologically characterized by the deposition of amyloid-β (Aβ) and the development of neurofibrillary tangles in the brain. Over time, an increased accumulation of these proteins leads to an adverse change in brain structure and elevated vascular injury, which are determined through quantifiable brain atrophy and WMH, respectively.

Typically, the progressive pre-symptomatic stage of AD lasts between 10 and 20 years. This could be the reason why 75% of people with dementia remain undiagnosed for prolonged periods. The availability of anti-amyloid therapies has increased the urgency for early detection of people with AD or mild cognitive impairment (MCI), as early diagnosis amplifies treatment benefits.

AD diagnosis is based on cognitive testing, which assesses the Aβ and phosphorylated tau burden using fluid biomarkers, positron emission tomography (PET), and magnetic resonance imaging (MRI). Clinicians can determine the changes in brain structure and integrity from multi-contrast MRI. These imaging indicators include generalized and hippocampal atrophy, which helps physicians understand the course of disease progression and cognitive decline.

Although neuroimaging immensely helps in AD and MCI diagnosis and management, its limited local and global accessibility contributes to its underdiagnosis. A previous study has highlighted the development of a portable LF-MRI, which could effectively increase accessibility and improve the diagnosis of different neurodegenerative diseases. This study highlighted the safety profile and low-cost, point-of-care scanning potential of LF-MRI. However, its reduced magnetic field strength lowers the signal-to-noise ratio (SNR), affecting image resolution.

About the study

The current study addressed the aforementioned limitation of LF-MRI for AD and MCI diagnosis by developing machine learning tools that can automatically quantify brain morphometry and white matter lesions.

An imaging pipeline was established that helped quantify brain volumes. The refined super-resolution and contrast synthesis technique (LF-SynthSR) was optimized to elevate LF image resolution in subsequent segmentation (SynthSeg). For example, hippocampal volumes derived from LF-MRI showed close agreement with high-field MRI counterparts, with an Absolute Symmetrized Percent Difference (ASPD) of 2.8% and a Dice similarity coefficient of 0.87. This strategy helped establish the optimal LF acquisition parameters for accurate quantification. It enabled white matter hyperintensity (WMH) burden (WMH-SynthSeg) measurement using automated segmentation of WMH lesions from T2 fluid-attenuated inversion recovery (FLAIR) images acquired at LF. This study validated the LF-SynthSR, SynthSeg, and WMH-SynthSeg using a prospective cohort of patients diagnosed with MCI or AD.

To establish an imaging pipeline, participants from three cohorts were included to undergo MRI acquisition on a portable, low-field 0.064 T MRI with a high-field, conventional scan at a field strength of 1.5–3 T. The first cohort contained twenty healthy individuals (10 males and 10 females) without a history of neurological disease or memory complaints.

The second cohort contained 23 participants (11 males and 12 females) having at least one vascular risk factor. However, none of the participants had any neurologic complaints or a prior history of memory disorder. The third cohort included 54 individuals (32 males and 22 females) with a diagnosis of MCI or AD. These participants underwent an LF-MRI imaging protocol that included T1w, T2w, and FLAIR sequences.

Study findings

Although the LF-MRI images did not have adequate resolution for automatic segmentation with high-field software analysis tools, they were first super-resolved (SR) to 1 mm isotropic T1-weighted (T1w) magnetization-prepared rapid gradient-echo (MP-RAGE)-like images. The study found that isotropic voxel sizes of ≤3 mm improved segmentation accuracy, producing ASPD values of less than 5% for hippocampal volumes. In addition, the accuracy of automated segmentation improved with the refinement of the LF-SynthSR v2 pipeline, enabling greater usability for low-field imaging applications.

In the first cohort, the accuracy of automated segmentation was assessed by comparing AD-relevant segmentation volumes of the hippocampus, lateral ventricle, and whole brain generated from the original LF-SynthSR and LF-SynthSR v2 against conventional high-field (HF) MRI acquired at 3 T.

An improvement in lateral ventricle volume accuracy was achieved by comparing LF-SynthSR v2 with LF-SynthSR v1. Image acquisition time ranged between 1:53 and 9:48 minutes, depending on voxel size and sequence. The study also found that isotropic voxel sizes of ≤3 mm improved segmentation accuracy, particularly in the low SNR regime of LF-MRI. The accuracy of brain morphometry was found to be affected by voxel size and geometry. Furthermore, LF-SynthSR v2 segmentation pipeline was validated against HF T1w MP-RAGE segmentations derived from the FreeSurfer segmentation tool ASEG.

WMH lesions due to axonal loss or cerebral small vessel disease were common among patients with cognitive impairment and were quantified using WMH-SynthSeg. The use of these findings on FLAIR as T2 hyperintense lesions and automated quantification of these lesions elevated the AD diagnosis and monitoring capacity of LF-MRI.

This study used machine learning to produce WMH lesion volumes (WMHv) from LF-FLAIR images using WMH-SynthSeg. This strategy enabled simultaneous segmentation of WMH T2 FLAIR lesions in addition to the prior brain morphometry. The WMH volumes correlated strongly with manual annotations and high-field imaging standards.

Based on WMHv generated by WMH-SynthSeg, the machine learning tool was validated as it was able to detect patients with MCI, AD, and those who are cognitively normal (CN).

Conclusions

The current study demonstrated that LF-MRI with machine learning tools can diagnose patients with AD or MCI. In the future, this device could also be assessed for its ability to detect neurodegenerative tauopathies and vascular dementia. Its portability, low cost, and automated analysis pipeline suggest significant potential for addressing diagnostic disparities globally.