NEC-Xpert: A Public Dataset of Neonatal Abdominal Radiographs with Ground Truth Labels, Multi-Reader Expert Annotations, and LLM-Assisted Interpretations

This dataset comprises 952 anonymized abdominal X-rays (AXRs) from 373 infants aged three months or under, all of whom were imaged at Great Ormond Street Hospital for Children due to clinical suspicion of necrotizing enterocolitis (NEC). The images were collected over a 15-year period as part of standard clinical care using routine pediatric imaging equipment. To be included, each case needed at least one usable X-ray and a clear, consensus-based final diagnosis; studies with inadequate image quality or diagnostic ambiguity were left out. The dataset provides the original PNG images alongside detailed structured annotations. A multidisciplinary panel of pediatric radiologists, surgeons, and neonatologists jointly established the reference standard for each case, categorizing patients as either NEC or non-NEC, and where NEC was confirmed, distinguishing between medical and surgical management. Each case was independently reviewed by three annotators using the same structured reporting template: a version of ChatGPT (v5.2) prompted via a custom NEC interpretation guide and a set of standardized prompt sequences; a junior pediatric radiologist with two years of relevant experience; and a senior pediatric radiologist with fifteen years in the field. All annotators documented their NEC classification (including an uncertain option), predicted management type for confirmed NEC cases, and noted the presence or absence of specific radiographic findings associated with Bell staging, such as bowel dilatation, pneumatosis intestinalis, portal venous gas, pneumoperitoneum, ascites, and a gasless abdomen. The dataset also includes the AI-generated reports used during a re-review phase, in which the human readers — without access to their original assessments — re-evaluated all cases with the LLM output available to them. All materials have been made publicly available to support a range of research applications, including the development and benchmarking of AI tools for neonatal imaging, exploration of human-AI interaction in radiology workflows, analysis of diagnostic variability, and use in pediatric radiology education. The included prompts and training materials are intended to allow researchers to replicate or extend the work across other AI architectures.