DeepSWI: Using Deep Learning to Enhance Susceptibility Contrast on T2*-Weighted MRI

Cerebral microbleeds represent a significant clinical concern, yet identifying them often requires specialized imaging protocols. Susceptibility-weighted imaging serves as the primary diagnostic standard for visualizing these small vascular lesions. However, the necessary phase information remains frequently unavailable during routine clinical examinations. This limitation restricts the ability of clinicians to accurately assess microbleed burden in many patients. Prior research has shown that T2*-weighted magnitude images provide some anatomical detail but lack the enhanced contrast of susceptibility-weighted sequences. That uncertainty drove the development of postprocessing techniques to bridge this diagnostic gap. No prior work had resolved how to reliably synthesize high-contrast images from standard magnitude data alone. This study addresses the need for accessible, high-quality neurovascular visualization in clinical settings.

The study aims to create synthetic susceptibility-weighted images from clinical T2*-weighted magnitude data using deep learning. This project seeks to overcome the lack of phase information in routine clinical magnetic resonance imaging. The researchers intend to evaluate the similarity of these synthetic images to conventional susceptibility-weighted standards. They also aim to determine if the generated images improve the detection of radiation-associated cerebral microbleeds. This work addresses the clinical challenge of accurately assessing microbleed burden without specialized acquisition protocols. The team motivated this investigation by the need for a postprocessing tool that enhances contrast in standard scans. No prior work had successfully demonstrated this level of synthetic contrast generation for clinical neurovascular assessment. This study provides a framework for improving diagnostic accuracy in patients where conventional susceptibility-weighted imaging remains unavailable.

The investigators conducted a retrospective analysis using a cohort of 145 adult participants. They partitioned the dataset into training, validation, and testing sets to optimize network performance. The team processed 3D gradient-echo scans to generate synthetic susceptibility-weighted outputs. They compared these synthetic results against original susceptibility-weighted and magnitude images using various quantitative metrics. Three blinded raters evaluated the quality and classification of the test-set images independently. The researchers applied Kruskall-Wallis and Wilcoxon signed-rank tests to determine statistical significance among the image groups. They utilized intraclass correlation to assess the consistency of ratings between the observers. This approach ensured a comprehensive evaluation of the model's ability to replicate gold-standard contrast.

The synthetic images achieved structural similarity index values of 0.972, 0.995, and 0.9864 for whole brain, vascular, and tissue regions. These values were statistically significantly higher than those observed between magnitude and original susceptibility-weighted images. Regarding microbleed detection, the synthetic model identified 67% of the lesions found on conventional susceptibility-weighted scans. In contrast, standard magnitude images only detected 36% of the same microbleeds. The overall quality of the synthetic images appeared comparable to the gold-standard sequences. The researchers observed fewer susceptibility-induced artifacts in the synthetic outputs than in the original images. Statistical analysis confirmed these improvements with p-values below 0.005. The findings demonstrate that the proposed network effectively increases contrast in the neurovasculature.

The authors propose that their deep learning model successfully enhances susceptibility contrast in neurovasculature. Their findings suggest that synthetic images provide a viable alternative when conventional phase data are missing. The researchers report that the generated outputs exhibit fewer artifacts compared to original susceptibility-weighted images. They claim that this approach improves the detection rate of radiation-associated microbleeds relative to standard magnitude scans. The study indicates that synthetic imaging maintains high structural similarity to gold-standard sequences. The team asserts that their method facilitates more precise estimation of lesion burden in clinical practice. These results imply that artificial intelligence can effectively augment existing magnetic resonance imaging protocols. The authors conclude that their framework offers a robust solution for enhancing diagnostic clarity without requiring additional acquisition time.