Dual-stage deep learning framework for neuroblastoma differentiation by integrating cell segmentation and multiscale modeling.

Background: As the most common extracranial solid tumor in children, accurate and consistent pathological diagnosis of neuroblastoma (NB) is critical for clinical decision-making. However, traditional methods relying on manual interpretation are constrained by tumor heterogeneity and inter-observer variability, necessitating the development of objective and quantitative intelligent auxiliary diagnostic technologies. Methods: This study introduces a two-stage deep learning framework with hybrid supervision, achieving precise classification of NB through collaborative optimization of cell-level segmentation and multiscale classification. In the first stage, the enhanced Medical Segment Anything Model (MedSAM) uses a cross-attention mechanism to achieve pixel-level tumor cell segmentation, attaining a Dice coefficient of 0.94. In the second stage, an optimized Swin Transformer (ST) is employed to construct the classification network, enabling full slice analysis via a confidence voting strategy. Results: On an independent dataset comprising 185 whole-slide images from 185 patients, the model attained an overall area under the receiver operating characteristic curve (AUC) of 0.864 (95% confidence interval (CI): 0.747 to 0.952), significantly outperforming existing methods (8.7% higher than the second-best model, ResNeXt). Among the three NB subtypes (undifferentiated, poorly differentiated, and differentiating), the recognition accuracy is the highest for the poorly differentiated subtype, which accounts for the largest proportion in clinical practice. Conclusion: This approach effectively addresses issues related to the tumor microenvironment interference and small-sample generalization through a cascaded feature extraction and decision-making mechanism, providing a robust intelligent auxiliary tool for NB pathological diagnosis.