A deep learning-enabled single breath-hold abbreviated MRI protocol may offer a faster, gadolinium-free route to hepatocellular carcinoma diagnosis while retaining information usually obtained from a full liver MRI examination. Hepatocellular carcinoma remains associated with delayed diagnosis, and ultrasound has limited sensitivity for small lesions. Conventional complete MRI offers higher sensitivity, but routine use faces constraints linked to scan duration and cost. A 2026 analysis published in Radiology: Artificial Intelligence assessed a protocol that acquires only a precontrast T1-weighted image during a single breath-hold and then uses generative artificial intelligence to create the remaining diagnostic sequences. The protocol aims to combine rapid acquisition, avoidance of gadolinium-based contrast agents and preservation of full-sequence imaging information.
A Single Breath-Hold Builds a Full MRI Set
The deep learning-enabled single breath-hold abbreviated MRI protocol uses one acquired sequence as its starting point. Patients at high risk of hepatocellular carcinoma undergo a rapid liver MRI scan of about 15 seconds to obtain a precontrast T1-weighted image. A generative artificial intelligence model then creates synthetic T2-weighted imaging, diffusion-weighted imaging, apparent diffusion coefficient maps and postcontrast T1-weighted sequences in arterial, portal venous and delayed phases. The acquired image and synthetic sequences together form the diagnostic image set.
The protocol addresses trade-offs seen in existing abbreviated MRI approaches. Non-contrast abbreviated MRI avoids gadolinium-based contrast agents but lacks some key imaging features used for definitive hepatocellular carcinoma diagnosis, including arterial phase hyperenhancement and washout. Contrast-enhanced abbreviated MRI requires gadolinium-based contrast agents, involves longer room times and often leaves out ancillary sequences such as T2-weighted and diffusion-weighted imaging. The new approach seeks to provide the diagnostic information of complete MRI while limiting acquisition to a single breath-hold precontrast scan.
The dataset included 1,008 patients at high risk of hepatocellular carcinoma from four institutions. The datasets included internal prospective, external retrospective and independent prospective cohorts collected between January 2019 and January 2025. All included patients had conventional complete MRI examinations, which provided the comparison image set and the input data for model development and testing.
Diffusion Model Selected for Multisequence Synthesis
Four generative models were trained to synthesise full MRI sequences from the precontrast T1-weighted input. The models included a conditional generative adversarial network, a Pixel2Pixel generative adversarial network, a Wasserstein generative adversarial network and a diffusion-based model named Li-DiffNet. Each model used six structurally identical submodules to generate the target sequences independently from the same precontrast input.
Li-DiffNet produced the strongest synthetic image quality across the model comparison and became the backbone of the protocol. Quantitative image quality assessment favoured Li-DiffNet across similarity, error and perceptual quality metrics, including structural similarity, mutual information, peak signal-to-noise ratio, mean absolute error, mean squared error and learned perceptual image patch similarity. Visual assessment also favoured the diffusion-based model for image fidelity, signal-to-noise ratio and lesion conspicuity.
Must Read: CT and MRI Signal HCC Subtypes for Individualised Care
The resulting protocol generated images that closely resembled conventional complete MRI in the reader assessment. Subjective image quality for the abbreviated approach remained non-inferior to conventional complete MRI across internal, external and prospective test sets. Mean image quality scores were close between the two methods, with conventional complete MRI scoring 4.18 to 4.19 and the deep learning-enabled protocol scoring 4.07 to 4.16 on a 5-point scale. Radiologists frequently misidentified the synthetic image sets as real complete MRI, with fooling rates above half across all test sets. Lesion size measurements also showed good agreement with conventional complete MRI.
Diagnostic Performance Remains Close to Complete MRI
The protocol achieved non-inferior diagnostic performance compared with conventional complete MRI at both patient and lesion levels. At the patient level, sensitivity for the deep learning-enabled protocol ranged from 81.2% to 88.7% across the internal, external and prospective test sets, while specificity ranged from 91.6% to 93.1%. Conventional complete MRI achieved sensitivity from 88.4% to 92.5% and specificity from 94.1% to 95.2% across the same test sets.
At the lesion level, sensitivity for the deep learning-enabled protocol reached 86.2% in the internal test set, 84.0% in the external test set and 77.9% in the prospective test set. Conventional complete MRI achieved 89.2%, 88.4% and 84.4% respectively. Non-inferiority testing confirmed that the abbreviated protocol remained within the prespecified diagnostic performance margin across reported metrics. Compared with a reference non-contrast abbreviated MRI protocol using precontrast T1-weighted, T2-weighted and diffusion-weighted imaging, the deep learning-enabled protocol demonstrated higher lesion-level sensitivity.
Failure analysis identified smaller lesions as more difficult to detect with both conventional complete MRI and the deep learning-enabled approach. Lesions missed by the abbreviated protocol but detected by conventional complete MRI typically showed arterial phase hyperenhancement or washout on conventional complete MRI but not on the synthetic image set. Subjective image quality scores were similar between false-negative and true-positive groups, indicating that the sensitivity gap related more to depiction of key enhancement features than to overall image quality.
The deep learning-enabled single breath-hold abbreviated MRI protocol combines rapid acquisition, gadolinium avoidance and synthetic full-sequence imaging for hepatocellular carcinoma diagnosis. Li-DiffNet generated the best-performing multisequence image set and supported diagnostic performance close to conventional complete MRI across internal, external and prospective testing. The protocol also showed higher lesion-level sensitivity than the reference non-contrast abbreviated MRI approach. Important boundaries remain, including unavailable matched ultrasound data, diagnostic rather than true surveillance testing, a proof-of-concept sample size and the need for real-world workflow validation and regulatory clearance before clinical implementation.
Source: Radiology: Artificial Intelligence
Image Credit: iStock
References:
Zhang Y, Geng Z, Qian X et al. (2026) Development and Validation of a Deep Learning-enabled Single Breath-hold Abbreviated MRI Protocol for Hepatocellular Carcinoma Diagnosis. Radiology: Artificial Intelligence [preproduction]: e250914.
