Deep learning enables free-breathing MRI, reducing sedation and scan time in children

A deep-learning-powered cardiac MRI technique has allowed children to breathe normally throughout their scans while substantially reducing examination time, anesthesia exposure and the need for endotracheal intubation, according to a study published in Pediatric Radiology. The approach produced quantitative measurements of heart function that closely matched those obtained with conventional breath-hold imaging, suggesting that a procedure traditionally difficult for young patients could become faster, less invasive and easier to tolerate. The findings are especially relevant for infants, medically fragile children and patients with congenital heart disease, for whom prolonged anesthesia and mechanical ventilation can add significant clinical complexity to an already demanding examination.

Cardiac magnetic resonance imaging, or CMR, is a powerful tool because it can show the heart’s anatomy, movement, blood flow and tissue characteristics without exposing patients to ionizing radiation. Yet the technology has a practical weakness: conventional cardiac cine imaging often requires a patient to hold their breath repeatedly while the scanner collects synchronized images across the cardiac cycle. Adults can usually follow these instructions, but young children may be unable to do so, and critically ill patients may not be physically capable of holding their breath. To prevent motion from blurring the images, hospitals frequently rely on sedation or general anesthesia. In some cases, that means placing a breathing tube and using muscle relaxation, extending the procedure beyond the scan itself.

The new method combines an unusually rapid acquisition strategy with a neural-network reconstruction system. Instead of collecting segmented images over several heartbeats while the patient performs breath-holds, the researchers used a highly accelerated cine sequence acquired during a single cardiac cycle, described as a “1RR” acquisition. The sequence captures the relevant information in one R–R interval—the period between two successive electrical activations of the heart—while the child continues breathing freely. Because less raw data are collected than in a conventional scan, the resulting signal is incomplete or “undersampled.” Deep-learning reconstruction is then used to infer and recover the missing image information, producing a clinically interpretable moving picture of the beating heart.

Deep-learning reconstruction does not mean that the algorithm is independently diagnosing disease. Rather, it functions as a sophisticated image-processing tool trained to recognize the structures and signal patterns expected in cardiac MRI data. The network reconstructs an image from abbreviated measurements while preserving boundaries between the blood pool and the myocardium, the muscular wall of the heart. That distinction matters because clinicians use cine MRI to trace the endocardial borders and calculate ventricular volumes, stroke volume and ejection fraction. If reconstruction were to smooth away anatomical edges or create artificial textures, the images might look attractive but yield unreliable clinical measurements. The study therefore assessed both visual image quality and numerical agreement with conventional imaging.

The investigators compared 150 children examined with the free-breathing, deep-learning approach with 150 sequential patients who underwent conventional segmented breath-hold cine imaging. The groups were evaluated for total CMR duration, whether anesthesia was used, how long anesthesia lasted and how quickly patients emerged from it. The difference in scanner time was substantial: examinations using the single-cycle free-breathing sequence lasted an average of 29.2 minutes, compared with 43.0 minutes for conventional imaging. That represents a reduction of nearly 14 minutes on average, or roughly one-third of the scan duration, although individual times varied in both groups.

The anesthesia findings were even more striking among children who still required sedation. Endotracheal intubation with muscle relaxation was used in 14 percent of patients undergoing the free-breathing deep-learning examination, compared with 90 percent of those receiving conventional breath-hold imaging. The study does not suggest that anesthesia became unnecessary for every child; some patients may still be too young, anxious or medically unstable to remain still in an MRI scanner. However, the results indicate that when sedation was needed, clinicians could often avoid the most invasive form of airway management. Average anesthesia duration fell from 98.3 minutes with conventional imaging to 70.8 minutes using the new sequence, while emergence time decreased from 27.1 minutes to 15 minutes.

Shortening anesthesia is clinically important because the risks of sedation do not end when image acquisition stops. Children must be monitored as they regain protective reflexes and normal breathing, and those with complex heart disease or other serious illness may require particularly careful recovery. Reducing the time spent anesthetized can simplify staffing, improve patient flow and potentially allow more examinations to be completed during a day. Avoiding intubation may also reduce the need for airway instrumentation and mechanical ventilation. The researchers emphasize an association between the imaging protocol and these reductions rather than claiming that the technology alone caused every difference, because the comparison involved sequential patient groups rather than a randomized trial.

A faster scan would have little value if it compromised diagnostic accuracy, but the researchers found no meaningful degradation in the reported contrast-to-noise performance. Using the contrast between the blood pool and myocardium, the measured value was 534.2 ± 165.1 with the free-breathing technique, compared with 491.8 ± 165.8 for conventional imaging. The reported statistical comparison was significant at P=0.03, yet the authors interpreted the overall result as maintaining diagnostic image quality rather than showing a clinically important loss of contrast. In MRI, contrast-to-noise ratio describes how clearly a structure can be distinguished from background variation; higher values generally make anatomical borders easier to assess, although clinical usefulness depends on more than a single numerical metric.

The researchers also examined whether the accelerated sequence could support reliable measurements of blood flow and ventricular performance. In a subgroup of 125 patients without significant valve regurgitation, measurements derived from free-breathing imaging showed strong correlations with flow measurements in the major vessels. The correlation between right ventricular cardiac index and main pulmonary artery flow was R=0.92, while the correlation between left ventricular cardiac index and aortic flow was R=0.84; both associations were statistically significant at P<0.001. Cardiac index adjusts cardiac output for body size, making it useful when comparing patients of different ages and sizes. The agreement suggests that the reconstructed cine images retained the temporal and spatial information needed to quantify how effectively the ventricles pump blood.

The findings may be particularly consequential for children weighing less than 35 kilograms, a group in which breath-holding and cooperation can be especially challenging. The source reports a subgroup analysis, although the detailed subgroup outcomes are not provided in the available article information. The broader results nevertheless point toward a workflow in which children can remain spontaneously breathing while the scanner acquires enough information to reconstruct a complete cardiac motion sequence. This could also help hospitals manage patients who are too sick for prolonged ventilation or who face repeated imaging during treatment. For congenital heart disease, where serial measurements are often needed to monitor anatomy and surgical repair, even modest reductions in anesthesia exposure can accumulate over time.

The technique is not a universal replacement for established cardiac MRI protocols. Deep-learning reconstruction must be validated across scanner models, field strengths, patient populations and disease patterns, and clinicians must remain alert to artifacts or failure modes that may not be obvious in a reconstructed image. The study was conducted under approval from the University of Pittsburgh Institutional Review Board, with informed consent waived, and included patients imaged in clinical practice rather than a randomized prospective experiment. The authors report no conflicts of interest, although two investigators were affiliated with GE HealthCare, which supplied technical expertise. The article also states that no datasets were generated or analyzed during the current study, a limitation for readers seeking independent access to the underlying data.

Even with those caveats, the work illustrates how artificial intelligence can change medical imaging without replacing the radiologist or cardiologist. The algorithm’s most immediate benefit is operational: it makes acquisition fast enough that ordinary breathing is less disruptive, reducing the need to immobilize a child through prolonged anesthesia. The clinical payoff follows from that engineering improvement. In the study, the images remained suitable for diagnosis and quantitative assessment while the time in the scanner, time under anesthesia and time recovering from anesthesia all fell. If future studies confirm the performance in younger children, infants and patients with more severe disease, free-breathing cardiac MRI could turn one of pediatric imaging’s most difficult balancing acts—capturing a moving heart from a moving child—into a considerably simpler procedure.