Deep Learning–based Multiple Arterial Phase MRI: A Step toward Improved Hepatocellular Carcinoma Diagnosis

See also the article by Liu and Zheng et al.

Mengchao Zhang, MD, PhD, is a professor in the Department of Radiology at China-Japan Union Hospital of Jilin University. His research focuses on artificial intelligence, advanced MRI techniques, and their clinical applications, particularly in the field of abdominal oncology.

Liver MRI plays a central role in the diagnosis, staging, treatment response assessment, and follow-up of hepatocellular carcinoma (1). According to the Liver Imaging Reporting and Data System, high-quality late arterial phase imaging is essential for hepatocellular carcinoma diagnosis—specifically, for accurate identification of nonrim arterial phase hyperenhancement, a key feature for classifying a lesion as LR-5 (definite hepatocellular carcinoma) (2). The late arterial phase is defined as the stage when the hepatic artery is fully enhanced, the portal vein is moderately enhanced, and the hepatic vein is not yet enhanced. However, obtaining optimally timed, diagnostically adequate late arterial phase images remains challenging in clinical practice, particularly with gadoxetate disodium, a hepatobiliary contrast agent (HBA).

Although gadoxetate disodium offers the dual benefit of dynamic imaging and hepatocyte-specific uptake, its standard dose contains only one-quarter the gadolinium of extracellular contrast agents (ECAs), resulting in weaker arterial enhancement and a narrower late arterial phase window (3). More critically, transient severe motion artifacts occur for approximately 8%–20% of patients due to acute transient dyspnea after gadoxetate injection, which severely compromises accurate capture of the late arterial phase and image quality (4). In clinical practice, suboptimal arterial phase images often lead to contrast agent changes, delayed examinations, or repeat scans, increasing costs and potentially delaying timely treatment. Improving the success rate and image quality of arterial phase imaging has therefore remained a key priority in this field.

Recent study findings suggest that multiple arterial phase acquisition can help address these challenges (5). The core concept is to acquire multiple arterial phases within a single breath hold, covering the entire arterial window and thereby increasing the probability of capturing an optimal late arterial phase. With continued advances in fast imaging techniques—from parallel imaging and view-sharing to compressed sensing and their combinations—the technical approaches for multiple arterial phase MRI have continuously evolved (6,7).

Deep learning (DL) has recently emerged as a research hotspot in this area. DL reconstruction algorithms, trained on large datasets, can reconstruct high-fidelity images from highly undersampled k-space data while suppressing aliasing artifacts and preserving anatomic details. Unlike conventional compressed sensing, DL does not require complex iterative reconstruction and can achieve comparable or better image quality within shorter reconstruction times. At abdominal MRI, DL has been applied to sequences such as half-Fourier acquisition single-shot turbo spin echo, diffusion-weighted imaging, and volumetric interpolated breath-hold examination (6), demonstrating value in reducing acquisition time, improving image sharpness, and enhancing lesion conspicuity. Combining DL with multiple arterial phase acquisition holds promise for achieving truly high-quality fast imaging by maintaining temporal resolution while overcoming potential spatial resolution loss associated with previous acceleration techniques.

The study by Liu and Zheng et al (8) published in Radiology: Imaging Cancer represents a clinical exploration of DL-based ultrafast multiple arterial phase acquisition in the context of this technical evolution. In this prospective study of 236 participants with suspected hepatocellular carcinoma according to prior CT or US examinations (128 who underwent ECA-enhanced MRI; 108 who underwent HBA-enhanced MRI), the authors used a DL-based ultrafast dynamic acquisition framework to obtain six consecutive arterial phases within a single breath hold, achieving a temporal resolution of 2.5 seconds, with DL reconstruction preserving spatial resolution comparable to the conventional single arterial phase imaging. The authors systematically evaluated the late arterial phase capture rate, image quality, and diagnostic performance and compared them with those for single arterial phase imaging.

The results showed that DL-based multiple arterial phase acquisition achieved significantly higher late arterial phase capture rates than single arterial phase imaging—98% versus 81% for ECAs (P = .001) and 98% versus 85% for HBAs (P = .04). Image quality was also superior in the multiple arterial phase group across overall quality, motion artifacts, and lesion edge sharpness. Importantly, this improvement in image quality translated into better diagnostic performance. For ECAs, the lesion detection rate increased from 90% to 98%, arterial phase hyperenhancement detection increased from 88% to 96%, and the hepatocellular carcinoma diagnosis rate increased from 81% to 96%. For HBAs, the lesion detection rate increased from 88% to 97%, and the arterial phase hyperenhancement detection rate increased from 86% to 97%. Notably, detection rates for small lesions (≤2 cm) improved in both groups—a finding with clinical implications for early hepatocellular carcinoma screening.

These advantages were also confirmed in the cirrhosis subgroup. Patients with cirrhosis, who are a high-risk population for hepatocellular carcinoma, often present greater challenges in arterial phase imaging (9,10). In the ECA cirrhosis subgroup, multiple arterial phase acquisition achieved a late arterial phase capture rate of 100%, whereas that achieved by single arterial phase imaging was 78% (P = .005). In the HBA cirrhosis subgroup, the multiple arterial phase acquisition capture rate was 97%, whereas that for single arterial phase imaging was 79% (P = .06). DL-based multiple arterial phase acquisition demonstrated consistent clinical value in this most challenging patient population.

In addition, DL-based multiple arterial phase acquisition offered the potential for hepatic artery evaluation. Early arterial phases clearly depicted hepatic artery branches, whereas late arterial phases best showed tumor enhancement, allowing a single scan to meet both diagnostic and preoperative vascular assessment needs.

A notable strength of this study is its careful attribution of technical gains. Through a self-controlled analysis in the supplemental material, the authors showed that DL reconstruction alone did not provide a statistically significant improvement in single arterial phase imaging. This finding suggests that the diagnostic improvement stems primarily from the multiphasic acquisition strategy itself, with DL serving as an enabler that makes high-quality multiphasic acquisition feasible. This distinction is important when evaluating new technologies.

Several questions merit further consideration. The clinical implementation of six consecutive arterial phases raises practical questions about reading efficiency and workflow integration—factors not addressed in this study but critical for real-world adoption. Furthermore, caution is required regarding the generalizability of these findings because the study was conducted at a single institution using a specific 3-T scanner and DL reconstruction algorithm. Whether similar results can be achieved across different field strengths, vendor platforms, and patient populations remains to be demonstrated. These considerations do not diminish the study’s contributions but rather highlight the questions that naturally arise as new technologies move toward broader clinical implementation.

In conclusion, DL-based ultrafast multiphasic arterial phase MRI increases late arterial phase capture rates to over 98%, while improving image quality and diagnostic performance, and offers the possibility of assessing hepatic arterial anatomy within a single scan. This technology substantially reduces uncertainty at a critical step in hepatocellular carcinoma diagnosis, providing a clinically feasible option for improving arterial phase imaging success—particularly in HBA-enhanced MRI. However, this represents only one step in the evolution toward high-quality hepatocellular carcinoma diagnosis. The ultimate clinical value of this technology will depend on reproducible results in multicenter settings, proven impact on patient outcomes, and favorable cost-effectiveness analysis.

Disclosures of conflicts of interest

Please see ICMJE form(s) for author conflicts of interest. These have been provided as supplemental materials.