New Web Tool Speeds and Standardizes Pediatric Hodgkin Lymphoma Staging

Hodgkin lymphoma is one of the most commonly diagnosed malignancies in children and adolescents, and getting its stage right is one of the most consequential decisions in pediatric oncology. The Ann Arbor classification, first introduced in 1971 and still recommended by the National Comprehensive Cancer Network, determines whether a child receives limited-field therapy or a more aggressive combined-modality regimen. Yet assigning that stage requires a radiologist or nuclear medicine physician to systematically evaluate whole-body 18F-FDG PET scans for nodal involvement above and below the diaphragm, splenic disease, bulky tumors, and extranodal spread — a task that is both time-consuming and notoriously reader-dependent. Now, a team at Stanford University has shown that a transparent, web-based computer-aided staging tool can cut staging time significantly and, crucially, that the benefit is greatest for the least experienced readers.

The study, published in Frontiers in Nuclear Medicine, analyzed 18F-FDG PET/MRI scans from 56 pediatric patients with histologically confirmed Hodgkin lymphoma treated at Lucile Packard Children’s Hospital between February 2023 and December 2025. The cohort included 30 males and 26 females with a mean age of 15.6 years, ranging from 7 to 24. Under the expert consensus reference standard, five patients had stage I disease, 20 had stage II, 12 had stage III, and 19 had stage IV. Bulky disease was present in 12 patients and extranodal involvement in 18. PET/MRI was chosen as the imaging modality because it reduces radiation exposure in children while maintaining high diagnostic accuracy for whole-body staging — an important consideration in a population where cumulative dose matters.

Seven readers participated, spanning the full spectrum of training: one second-year radiology resident with no prior PET experience, four board-certified pediatric radiologists with 2 to 15 years of experience interpreting 90 to 1,500 PET examinations per year, and two board-certified nuclear medicine physicians with 7 to 9 years of experience interpreting 2,000 to 4,500 PET cases annually. Each reader staged the scans twice — once by traditional visual inspection and once with the new tool — with the 56 scans divided into two groups with matched Ann Arbor stages. An additional nuclear medicine physician and pediatric radiologist jointly established the reference standard, and their interpretations showed no discrepancies with the clinical reports.

The tool itself is architecturally distinct from most artificial intelligence systems in medical imaging. Rather than a deep learning model that predicts a stage from pixel data, it combines a rule-based decision tree reflecting the Ann Arbor criteria with an interactive web interface implemented primarily in JavaScript. After identifying lesions on the PET/MRI scan, the reader categorizes tumor sites as nodal or extranodal and assigns them to their anatomical distribution. The decision tree then evaluates staging criteria in descending order — beginning with stage IV, followed by III, II, and I — making a binary decision at each step based on imaging-defined tumor extent, such as the distribution of disease relative to the diaphragm. Users answer six structured questions covering lymph node involvement above and below the diaphragm, splenic involvement, and extranodal disease, with extent recorded as none, a single nodal station, or multiple stations, plus optional bulky disease designation. This structured approach makes staging deterministic and reproducible.

The results were striking. Mean staging time dropped from 13.87 minutes without the tool to 10.8 minutes with it, a statistically significant reduction (p = 0.001). But the headline finding was the inverse relationship between reader experience and time saved. The resident’s staging time fell from 21.5 to 17.63 minutes, the radiologists’ from 18.03 to 13.63 minutes, while the nuclear medicine physicians — who already stage in roughly 1.4 minutes — showed no significant change. Linear regression confirmed the pattern quantitatively: a significant inverse association between annual PET interpretation volume and time saved (β = −0.812, p < 0.001, R² = 0.66), meaning the tool delivers roughly two-thirds of its variance in benefit to those who need it most. The largest relative time reductions were seen for stage I and stage IV disease, at 36.13% and 36.30% respectively.

Accuracy improved as well, and again the gains concentrated among junior readers. Agreement with the reference standard, measured by Cohen’s kappa, rose from 0.433 to 0.789 for the resident — whose staging errors dropped from 11 to 4 — and from 0.476 to 0.946 for Radiologist 1, whose discrepancies fell from 10 to a single error. The nuclear medicine physicians, whose baseline kappa values were already 0.836 and 0.834, improved to 0.946 and 0.947 with just one error each. On the Landis and Koch scale, these shifts move junior readers from moderate agreement to substantial or almost perfect agreement with expert consensus. The tool reduced discrepancies for the resident and less experienced radiologists but offered limited added value for expert readers already performing at a high level.

The clinical stakes of this variability are not abstract. Prior studies cited by the authors reported discordance rates of 11.3% in disease staging and 7.8% in risk stratification between a multidisciplinary clinic and an independent expert team, with up-staging in all discordant cases leading to overtreatment. Another observer study found only 56% to 61% agreement between eleven nuclear medicine physicians and expert readers on baseline PET staging. In pediatric Hodgkin lymphoma, where treatment intensity is calibrated to stage, such variability directly risks both over- and undertreatment. A structured staging workflow that standardizes the process could be particularly valuable in settings with limited subspecialty expertise, supporting more consistent decision-making across institutions.

The tool’s transparent, rule-based design also addresses a persistent barrier to clinical AI adoption. Deep learning models in medical imaging have repeatedly been shown to rely on non-causal or confounding features — one pneumonia detection model lost performance on external datasets because it had learned imaging device characteristics rather than disease-specific features, and COVID-19 detection models have depended on spurious correlations unrelated to pathology. Saliency maps, the standard explanation method, only indicate where a model focuses without explaining what imaging features drive the prediction. By contrast, the Stanford tool follows a white-box approach in which each decision step is explicitly defined by established clinical criteria, allowing direct verification of the decision-making process and ensuring staging outcomes derive from clinically relevant inputs. The reader remains in control throughout, a deliberate contrast to black-box systems that exclude clinicians from the decision pathway.

The study has limitations worth noting. It was single-center and retrospective, with a modest cohort of 56 patients, and the readers staged the same scans under both conditions rather than in a fully blinded parallel design. The tool was tested on pediatric patients, though the authors suggest it could be applied to adults in the future. It also requires the reader to identify lesions manually — the tool standardizes the staging logic, not the lesion detection step, which remains the most time-consuming component of PET interpretation. Whether the time savings translate to busy clinical workflows with competing demands, or to centers where PET/MRI is less available, remains to be demonstrated in prospective multicenter studies.

Still, the implications are compelling. The authors report that FDG PET utilization at major U.S. cancer centers rose from 5.9 to 6.7 studies per 1,000 persons between 2011 and 2020, while the number of radiologists per 100,000 population remained essentially flat and the pediatric radiology workforce declined. A 20.2% misdiagnosis rate has been reported in on-call pediatric radiology reporting. Against that backdrop, a freely accessible web resource that raises junior readers’ staging accuracy to near-expert levels while cutting staging time by up to a third could meaningfully narrow the gap between imaging volume and reader capacity. The team has released the tool openly at mike-barrow.github.io/wbtdemo, inviting clinicians and researchers worldwide to test whether structured, clinician-guided decision support can do what black-box prediction has struggled to achieve: make expert-level staging reproducible everywhere.

Subject of Research: Computer-aided Ann Arbor staging of pediatric Hodgkin lymphoma

Article Title: Impact of reader experience on the time-saving potential of a novel computer-aided tool for Ann Arbor staging of pediatric hodgkin lymphoma

Article References: Vasyliv, I., Barrow, M. J., Baratto, L., Kim, H. G., von Krüchten, V. R., Lokesha, Y. U., Sarrami, A. H., Moeremans, M., Guja, K. E., Hawk, K. E., & Daldrup-Link, H. E. (2026). Impact of reader experience on the time-saving potential of a novel computer-aided tool for Ann Arbor staging of pediatric hodgkin lymphoma. Frontiers in Nuclear Medicine, 6, Article 1843299. https://doi.org/10.3389/fnume.2026.1843299

Image Credits: AI Generated

DOI: 10.3389/fnume.2026.1843299

Keywords: pediatric Hodgkin lymphoma, Ann Arbor staging, 18F-FDG PET/MRI, computer-aided diagnosis, radiology, nuclear medicine, artificial intelligence, Impact, reader, experience, time-saving, potential