Deep Learning Cuts Stroke MRI Acquisition Time

Faster MRI acquisition could support more timely assessment of patients with acute ischaemic stroke without sacrificing the image quality needed for diagnosis. A retrospective study in Emergency Radiology evaluated compressed sensing combined with deep learning reconstruction in 69 patients examined on a 3.0T system in China. The approach was applied to three structural brain sequences commonly used alongside diffusion imaging to assess tissue changes and exclude conditions that can resemble stroke. Compared with conventional acquisition, the accelerated reconstructions substantially reduced scan duration. Moderate acceleration produced the strongest balance between speed, overall quality and lesion clarity, while the fastest setting showed some loss of detail. 

Faster Structural Imaging in Acute Stroke 

Acute ischaemic stroke requires rapid imaging because diagnosis and treatment decisions are time-sensitive. Diffusion-weighted imaging is central to detecting early ischaemic lesions, but structural MRI sequences also contribute to the assessment of alternative causes of neurological symptoms. T1-weighted, T2-weighted and fluid-attenuated inversion recovery imaging can help identify tumours, encephalitis, metabolic disorders and other stroke mimics. These sequences may be abbreviated or omitted in urgent protocols when their acquisition extends examination time. 

Must Read: AI Supports Brain Tumour Detection on MRI 

The evaluated method combines compressed sensing, which reduces the amount of acquired data, with a deep learning network that reconstructs the undersampled images. Conventional full-sampling images served as the reference. The same datasets were retrospectively reconstructed at three acceleration levels, designated R3, R4 and R5. The model used multiple reconstruction stages to improve the image while maintaining consistency with the original signal. 

The combined acquisition time for the three sequences fell from just over six minutes with conventional imaging to about two minutes at the lowest acceleration level. It decreased further at the two higher levels, reaching around one and a half minutes with R4 and slightly over one minute with R5. The reductions were seen across all three sequences, with the longest conventional sequence showing the greatest absolute time saving. These results indicate that structural brain imaging can be compressed into a much shorter examination while retaining clinically usable images. 

Moderate Acceleration Preserves Image Quality 

Two radiologists assessed overall image quality, visible noise and structural clarity on a five-point scale. Quantitative analysis also measured signal-to-noise and contrast-to-noise performance in lesions and surrounding brain tissue. Agreement between the readers was excellent across the evaluated sequences, supporting the consistency of the qualitative assessments. 

All accelerated reconstructions achieved higher signal-to-noise and contrast-to-noise measurements than conventional imaging. The qualitative findings differed according to the acceleration level. R3 and R4 produced better overall quality and lower perceived noise than full-sampling images across the evaluated structural brain sequences. Lesion clarity was largely maintained at these settings, although clarity for one sequence was lower with R4 than with the conventional reference. 

The fastest reconstruction showed a less favourable balance. R5 retained overall quality comparable with conventional imaging and received better noise scores, but structural clarity was lower across the three sequences. This suggests that further acceleration can continue to reduce examination time while weakening the visibility of fine detail. 

Among the tested settings, R4 offered a notable reduction in acquisition time without a significant difference in diagnostic accuracy from conventional imaging. R3 also produced strong image quality but required a longer examination than R4. The results therefore support moderate rather than maximal acceleration when both speed and clarity are priorities in emergency brain MRI. 

Diagnostic Performance and Study Limits 

The accelerated T1-weighted and T2-weighted protocol was assessed for its ability to exclude stroke mimics using the final clinical discharge diagnosis as the reference. The protocol achieved sensitivity and specificity above 90%, with a negative predictive value of about 95%. Diagnostic accuracy with R4 did not differ significantly from conventional imaging. These results support the use of accelerated structural sequences as part of an assessment intended to distinguish acute ischaemic stroke from other conditions with similar presentations. 

The patient group included small lacunar infarcts as well as medium and large territory infarcts. Some patients also had large-vessel occlusion. However, the cohort was not large enough for reliable subgroup comparisons based on infarct size or vessel status. The results therefore do not establish whether reconstruction performance varies across these clinical categories. 

Several other limitations restrict wider application. The analysis was retrospective and came from one centre using a single 3.0T scanner. Performance on other systems, including lower-field scanners, was not evaluated. The reconstruction model also depends on the data used for training, which may not represent every pathological appearance. 

The evaluation covered structural sequences only. Diffusion-weighted imaging and susceptibility-weighted imaging, both important components of acute stroke protocols, were not included. Standardised magnetic resonance angiography was also unavailable for all patients. The results therefore apply to the specific morphological sequences tested rather than to a complete MRI stroke pathway. Broader evaluation would be needed across multiple centres, scanner types and full emergency protocols. 

Compressed sensing deep learning reconstruction substantially shortened structural brain MRI acquisition in patients with acute ischaemic stroke. Moderate acceleration provided the best overall balance, improving image quality and noise performance while largely preserving lesion clarity and maintaining diagnostic accuracy comparable with conventional imaging. The fastest setting reduced the examination further but produced lower clarity. The results support the feasibility of faster T1-weighted, T2-weighted and fluid-attenuated inversion recovery imaging in emergency assessment. Their wider use remains limited by the small single-centre cohort, the single scanner platform and the absence of several core stroke sequences from the evaluation. 

Source: Emergency Radiology 

Image Credit: iStock