Related Experiment Videos
Dual-stage deep learning framework for neuroblastoma differentiation by integrating cell segmentation and multiscale
Jieni Xiong1, Zhu Zhu2,3, Weizhong Gu4
1Surgical Oncology, Children's Hospital, Zhejiang University School of Medicine, National Clinical Research Center for Children and Adolescents' Health and Diseases, Hangzhou, Zhejiang, China.
World Journal of Pediatric Surgery
|July 9, 2026
Summary
A novel deep learning framework precisely diagnoses neuroblastoma (NB), a common childhood cancer. This AI tool enhances pathological diagnosis by overcoming limitations of manual interpretation, improving accuracy for critical clinical decisions.
Area of Science:
- Computational pathology
- Artificial intelligence in oncology
- Pediatric oncology
Background:
- Accurate pathological diagnosis of neuroblastoma (NB), the most common pediatric extracranial solid tumor, is crucial for treatment.
- Traditional manual diagnosis faces challenges due to tumor heterogeneity and inter-observer variability.
- Objective, quantitative intelligent diagnostic tools are needed to improve NB classification.
Purpose of the Study:
- To develop and validate a two-stage deep learning framework for precise neuroblastoma classification.
- To enhance diagnostic accuracy and consistency in neuroblastoma pathological analysis.
- To provide an objective auxiliary tool for clinical decision-making in neuroblastoma cases.
Main Methods:
- A two-stage deep learning framework utilizing hybrid supervision was developed.
- Stage 1: Enhanced Medical Segment Anything Model (MedSAM) with cross-attention for pixel-level tumor cell segmentation (Dice coefficient: 0.94).
- Stage 2: Optimized Swin Transformer (ST) for multiscale classification with a confidence voting strategy for full slice analysis.
Main Results:
- The model achieved an AUC of 0.864 on an independent dataset of 185 whole-slide images.
- Performance significantly outperformed existing methods, showing an 8.7% improvement over the second-best model.
- Highest recognition accuracy was observed for the poorly differentiated neuroblastoma subtype.
Conclusions:
- The deep learning framework effectively addresses tumor microenvironment interference and small-sample generalization.
- A cascaded feature extraction and decision-making mechanism provides a robust intelligent auxiliary tool.
- This AI-driven approach offers a significant advancement in neuroblastoma pathological diagnosis.