Construction and Validation of an Intelligent Ultrasound Diagnostic System for the Spectrum of Neuroblastoma in Children

Neuroblastic tumors (NTs) represent the most common extracranial solid tumors in childhood, with the vast majority of patients diagnosed with neuroblastoma (NB)-the subtype associated with the highest malignancy and poorest prognosis. These cases present significant challenges in clinical diagnosis and management, often leading to unfavorable overall outcomes. Histopathological examination remains the gold standard for definitive diagnosis and classification. However, this method is invasive, carries a risk of complications, and its diagnostic accuracy is subject to operator experience and biopsy sampling location. Although medical imaging allows for noninvasive tumor assessment, it primarily relies on subjective visual interpretation by physicians, resulting in limited accuracy and reproducibility in distinguishing between different NT subtypes.

Radiomics, an emerging artificial intelligence-based imaging analysis approach, enables high-throughput extraction, analysis, and quantification of imaging features through automated algorithms, uncovering vast amounts of subvisual information. It has demonstrated considerable promise in the differential diagnosis, treatment evaluation, and outcome prediction of tumors. Current radiomics research on neuroblastoma is still in its early stages, with most studies focusing on modalities such as computerized tomography(CT), magnetic resonance imaging(MRI), and Positron Emission Tomography-Computed Tomography(PET-CT), while ultrasound-based radiomics investigations remain unexplored.

Ultrasonography, owing to its unique advantages-including absence of ionizing radiation, real-time dynamic imaging, operational convenience, and low cost-has become the preferred imaging modality for pediatric tumor screening and follow-up. Consequently, integrating radiomics with ultrasonography to develop an intelligent diagnostic system capable of noninvasively and accurately assessing NTs holds significant clinical value and translational potential. Such a system would facilitate precise preoperative classification, patient risk stratification, and support for clinical decision-making.

This study aims to construct and validate an intelligent ultrasound diagnostic system for pediatric neuroblastic tumors based on ultrasound radiomics features, as follows:

• To build a prediction model for pediatric neuroblastic tumors (NTs) based on ultrasound images, achieving automated differential diagnosis of neuroblastoma (NB), ganglioneuroblastoma (GNB), and ganglioneuroma (GN).
• On the basis of pathological classification, integrate clinical pathological features to establish a multimodal fusion model. The focus is on identifying high-risk patients, predicting bone marrow metastasis, and estimating treatment outcomes, providing a reference basis for clinical decision-making.
• Integrate previous research results to construct a comprehensive intelligent ultrasound diagnostic system that integrates pathological classification, risk stratification, and prognosis assessment, achieving systematic and intelligent evaluation of NTs patients to assist in clinical risk stratification and individualized treatment decisions.

Inclusion criteria

• The diagnosis of NTs was confirmed by surgical resection or biopsy with histopathological examination, and the type was classified as NB, GNB or GN according to the INPC standard.
• Age ≤ 18 years old, with no gender restrictions.
• There are complete abdominal (or primary site) ultrasound images archived, in original DICOM or JPG format, with image quality meeting the analysis requirements.
• Complete clinical and pathological data relevant to the research purpose are available.

Exclusion criteria

• The patient has previously undergone surgical resection treatment in another hospital, but the tumor recurred or remained after the operation.
• Poor quality of ultrasound images: There are artifacts that seriously affect the identification of tumor contours or feature extraction, image blurring, or incomplete display of the lesion.
• Severe data deficiency: Key clinical pathological data or imaging data are missing, making it impossible to extract and analyze the required information.