Digital pathological classification of peripheral neuroblastic tumors across multi-continent cohorts
Authors:
Preprint
Abstract
Peripheral neuroblastic tumors (pNTs) are embryonal tumors of neural crest origin with highly variable clinical behaviors, ranging from spontaneous regression to aggressive and fatal disease. The International Neuroblastoma Pathology Classification (INPC) defines four diagnostic categories and stratifies tumors into Favorable Histology (FH) or Unfavorable Histology (UH) group based on age-linked assessment of tumor differentiation/maturation and the Mitosis-Karyorrhexis Index (MKI). Prognostic distinction according to INPC offers significant prediction of clinical outcomes, with patients harboring FH tumors demonstrating longer overall and event-free survival than those with UH tumor. Here, we assembled the largest multi-continental whole-slide histology imaging dataset of pNTs to date, comprising 2,030 tumors from the United States, Canada, Australia, New Zealand, Germany and Japan. Using this dataset, we developed NB-PathAssist, a computational pathology framework for automated INPC-based classification from digitized whole-slide images. NB-PathAssist integrates foundation-model-based feature extraction with an archetype-guided SummaryMixing architecture that compresses whole-slide morphology into representative histologic tokens for tumor-level classification. To automate MKI assessment, we manually annotated 14,049 nuclei, classifying them as mitotic-karyorrhectic (MK) tumor cells, non-MK tumor cells, or non-tumor cells. The resulting model was applied to phenotype cells in whole-slide images across 1,649 tissue sections. Across geographically and institutionally diverse cohorts, NB-PathAssist reproducibly assessed key diagnostic features and generated prognostic classifications concordant with reference INPC annotations. The average balanced accuracy for classifying tumor categories, grade of neuroblastic differentiation, MKI and prognostic groups achieved 0.84, 0.86, 0.89 and 0.96, respectively. We packaged the NB-PathAssist algorithms into an end-to-end online platform for scalable and standardized AI-assisted assessment of pNTs.Keywords
Funding Information
- U.S. Food and Drug Administration (75F40123C00103)
- Children's Oncology Group (U10CA180886)
- Children's Oncology Group (U24CA196173)
- Children's Oncology Group (U10CA180899)
- National Cancer Institute (1K99CA293249)
- Deutsche Forschungsgemeinschaft (493872418)
- Deutsche Forschungsgemeinschaft (413326622)
This publication has 26 references indexed in Scilit:
- Digital profiling of gene expression from histology images with linearized attentionNature Communications, 2024
- Integrative analysis of neuroblastoma by single-cell RNA sequencing identifies the NECTIN2-TIGIT axis as a target for immunotherapyCancer Cell, 2024
- Ganglioneuroblastoma intermixed: Clinicopathological implications of diagnosis at presentation and genomic correlationsPediatric Blood & Cancer, 2023
- Pathological prognosis classification of patients with neuroblastoma using computational pathology analysisComputers in Biology and Medicine, 2022
- Association of heterogeneous MYCN amplification with clinical features, biological characteristics and outcomes in neuroblastoma: A report from the Children's Oncology GroupEuropean Journal of Cancer, 2020
- Clinicopathological characteristics of ganglioneuroma and ganglioneuroblastoma: A report from the CCG and COGPediatric Blood & Cancer, 2009
- Computer-aided evaluation of neuroblastoma on whole-slide histology images: Classifying grade of neuroblastic differentiationPattern Recognition, 2009
- Revision of the International Neuroblastoma Pathology ClassificationCancer, 2003
- International neuroblastoma pathology classification for prognostic evaluation of patients with peripheral neuroblastic tumorsCancer, 2001
- The International Neuroblastoma Pathology Classification (the Shimada system)Cancer, 1999