In a prospective study published in BMC Medical Imaging, investigators from the Children’s Hospital of Zhejiang University School of Medicine and Zhejiang University report that a magnetic resonance technique called amide proton transfer imaging, analyzed with machine learning, can help distinguish high-risk from non-high-risk abdominal neuroblastoma before any tissue is taken. The work, led by co-first authors Wenqi Wang and Jiawei Liang with corresponding authors Hongxi Zhang and Yi Zhang, enrolled 121 consecutive children with suspected neuroblastoma and scanned them on a clinical 3-Tesla MRI scanner. Fifty-eight children, with a mean age of roughly three and a half years, ultimately formed the analysis cohort: 24 classified as non-high-risk and 34 as high-risk.

To understand why this matters, it helps to know how neuroblastoma risk is currently assigned. Clinicians combine a child’s age, the stage of the tumor, its histological appearance, and characteristic genetic abnormalities, most famously amplification of the MYCN oncogene, into frameworks such as the International Neuroblastoma Risk Group staging system. Each of those inputs typically requires biopsies, surgical sampling, bone marrow examinations, and a panel of molecular tests. For a toddler, every one of those procedures carries anesthesia risk, emotional toll, and delay. A noninvasive scan that could hint at biological aggressiveness before the first incision would reshape the entire diagnostic pathway, allowing families and physicians to plan with greater confidence.

Amide proton transfer imaging belongs to a family of methods known as chemical exchange saturation transfer, or CEST. The physics is elegant. Mobile proteins and peptides inside tissue carry amide groups whose hydrogen protons resonate at a frequency about 3.5 parts per million away from water. If a radiofrequency pulse selectively saturates those amide protons, they continuously exchange with the protons of surrounding water molecules, transferring the saturation like a relay baton and measurably dimming the water signal. The degree of that dimming, quantified as the chemical exchange saturation transfer ratio, becomes a proxy for the concentration and exchange behavior of mobile proteins in the tumor. Because malignant tissues often overproduce intracellular mobile proteins, aggressive tumors tend to light up differently from indolent ones, offering a molecular contrast mechanism that conventional T1-weighted and T2-weighted MRI cannot provide.

The Zhejiang team did not stop at a single CEST metric. They measured four related quantities from each tumor: the standard CESTR, a reference-normalized version called CESTR-nr, the inverse magnetization transfer ratio known as MTR-Rex, and a metric called AREX, short for apparent exchange-dependent relaxation. Each of these corrects for different confounding factors, such as direct water saturation and conventional magnetization transfer effects, that can contaminate raw CEST signals. After two experienced radiologists delineated initial tumor regions, a semi-automatic segmentation method refined the volumes of interest, and the four metrics were extracted from each child’s tumor. The researchers also assessed measurement reproducibility using intraclass correlation coefficients, an important safeguard when the entire diagnostic premise rests on numbers that must be trustworthy from scan to scan.

The analytical strategy was where the study became genuinely ambitious. Rather than relying on a single averaged number, the team built predictive models using seven widely used machine learning classifiers, including logistic regression, support vector machine, random forest, decision tree, k-nearest neighbor, naïve Bayes, and least absolute shrinkage and selection operator, commonly abbreviated LASSO. They compared single-metric models built from each APT parameter against models that combined multiple metrics, and they benchmarked everything against the simpler approach of using mean APT values alone. Performance was judged by the area under the receiver operating characteristic curve, or AUC, a standard measure of diagnostic discrimination in which 0.5 represents coin-flip accuracy and 1.0 represents perfect classification.

The results told a clear story. APT metrics separated risk groups more effectively than conventional MR images could. Single-metric models reached a maximum AUC of 0.81, a substantial leap over the mean APT values, which topped out at only 0.60. Among the four parameters, MTR-Rex proved the strongest performer, with its models achieving AUCs between 0.67 and 0.81, while models built on the other three metrics ranged from 0.62 to 0.72. The best result of all came from combining metrics: a LASSO-based model integrating MTR-Rex with either CESTR or AREX pushed the AUC to 0.83. In practical terms, that means the combined model correctly ranked a randomly chosen high-risk child above a randomly chosen non-high-risk child more than eight times out of ten, using nothing but molecular information harvested from a routine MRI session.

Why would MTR-Rex outperform its siblings? The metric is designed to isolate the exchange-dependent component of the saturation signal while suppressing contributions from semi-solid macromolecules and direct water saturation. In tumors, where cell density, necrosis, and protein content vary wildly from region to region, that cleaner isolation appears to preserve the biological information that actually tracks aggressiveness. The success of the multiparametric combination reinforces a principle increasingly recognized across radiology: no single number captures a tumor’s complexity, but a small, well-chosen panel of quantitative features, fed through an appropriate classifier, can approach the discriminative power of tissue itself.

The clinical implications extend beyond neuroblastoma. APT imaging is already under investigation in brain tumors, breast cancer, and head and neck malignancies, but applying it to abdominal tumors in small children poses distinct challenges, including respiratory motion, small anatomical targets, and the need for short scan times in patients who cannot cooperate or safely receive sedation for long. Demonstrating feasibility on a standard 3-Tesla clinical scanner, rather than a research-only system, is a meaningful step toward translation. The study also benefited from funding through China’s National Key Research and Development Program, the National Natural Science Foundation, and the Zhejiang Provincial Natural Science Foundation, with one co-author affiliated with Philips Healthcare, reflecting the growing partnership between academic imaging labs and scanner manufacturers.

Caution is still warranted. Fifty-eight analyzed cases is a modest sample, and the authors themselves frame the models as potential clinical aids rather than replacements for biopsy and genetic testing. The published version is an early-release article subject to final edits, and external validation in independent, multi-center cohorts will be essential before any risk-stratification decision rests on an APT-derived number. Nevertheless, the direction of travel is unmistakable. If larger studies confirm these findings, children with abdominal masses could one day undergo a single MRI that simultaneously maps anatomy and interrogates tumor chemistry, giving oncologists an early, noninvasive read on how dangerous a tumor is likely to be. For the youngest cancer patients, whose tumors are often discovered at their most treatable moment but whose bodies are least able to tolerate invasive diagnostics, that would be a genuinely transformative advance.

Subject of Research: Amide proton transfer MRI with machine learning for preoperative risk stratification of pediatric neuroblastoma

Article Title: Multiparametric analysis of amide proton transfer imaging for preoperative risk stratification of pediatric neuroblastoma

Article References: Wang, W., Liang, J., Jia, X., Wen, J., Ma, X., Chen, W., Wu, D., Lai, C., Zhang, H., & Zhang, Y. (2026). Multiparametric analysis of amide proton transfer imaging for preoperative risk stratification of pediatric neuroblastoma. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02815-0

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02815-0

Keywords: neuroblastoma, amide proton transfer, CEST MRI, risk stratification, machine learning, pediatric oncology, MTR-Rex, LASSO, 3T MRI, biomarkers, diagnostic imaging, predictive medicine