This is a multi-center, retrospective, observational study to develop and internally validate an artificial intelligence (AI) foundation model for hierarchical classification of central nervous system (CNS) tumors using approximately 20,000 hematoxylin and eosin (H\&E) whole-slide images (WSIs) collected at Huashan Hospital Fudan University and Shandong Provincial Hospital. Archived pathology slides and linked de-identified clinical, histopathological, and molecular diagnostic data from patients who underwent neurosurgical tumor resection or biopsy between January 1, 2010 and December 31, 2025 will be retrospectively analyzed. The study aims to train and evaluate weakly supervised multiple-instance learning models using pathology foundation models and conventional convolutional neural network feature extractors to predict tumor category, tumor family, terminal WHO 2021 CNS tumor diagnosis, and selected molecular alterations directly from routine H\&E slides. Internal model validation will be performed using patient-level training, validation, and hold-out test datasets. Secondary analyses include comparison of model architectures, virtual molecular profiling, interpretability analyses using attention heatmaps, and comparison of AI-assisted versus pathologist-only diagnostic performance on selected internal test cases.
Central nervous system tumors comprise a highly heterogeneous group of neoplasms with substantial diagnostic complexity. The WHO 2021 Classification of Tumors of the Central Nervous System integrates histology with molecular biomarkers, making accurate diagnosis increasingly dependent on molecular features such as IDH mutation, 1p/19q codeletion, H3 alterations, TERT promoter mutation, and other genomic or epigenomic markers. However, broad implementation of comprehensive molecular testing remains limited in many settings because of cost, turnaround time, technical complexity, and tissue constraints. This retrospective study will use archived formalin-fixed paraffin-embedded H\&E glass slides or existing digital WSIs from approximately 20,000 patients with primary or secondary CNS tumors treated at Huashan Hospital, Fudan University and Shandong Provincial Hospital. Slides will be digitized when necessary, de-identified, quality controlled, segmented for tissue regions, and divided into image patches. Patch-level features will be extracted using pretrained image encoders, including ResNet50, UNI, and CONCH, followed by weakly supervised multiple-instance learning aggregation methods such as attention-based MIL and CLAM. The primary objective is to develop and internally validate an AI model capable of hierarchical CNS tumor classification, including tumor category, tumor family, and terminal WHO 2021 diagnosis. Secondary objectives are to compare alternative model architectures, evaluate prediction performance for key molecular markers, assess model interpretability with attention mapping, and compare AI-only, pathologist-only, and AI-assisted diagnosis on an internal test subset. No intervention will be delivered to participants, and no clinical treatment decisions will be based on model outputs during this research stage. All data processing and model development will be conducted on secure in-hospital servers using de-identified data in accordance with institutional ethics approval and data protection procedures. ClinicalTrials.gov defines observational studies as studies in which investigators assess outcomes without assigning interventions, which matches this study design.
Study Type
OBSERVATIONAL
Enrollment
20,000
Hierarchical CNS tumor classification performance on the internal hold-out test set
Diagnostic performance of the final AI model for hierarchical classification of CNS tumors at the tumor category, tumor family, and terminal WHO 2021 diagnosis levels using de-identified H\&E whole-slide images. Performance metrics will include macro- and/or micro-area under the receiver operating characteristic curve (AUC), balanced accuracy, weighted F1 score, and Matthews correlation coefficient (MCC).
Time frame: Assessed at model evaluation after completion of training, up to Jul 2029
Comparative performance of alternative feature extractors and MIL aggregation methods
Comparison of model performance across feature extractors (ResNet50, UNI, CONCH) and aggregation methods (ABMIL, CLAM) on the internal validation and hold-out test datasets using AUC, balanced accuracy, sensitivity, specificity, weighted F1 score, and MCC.
Time frame: Up to Jul 2029
Prediction performance for selected molecular biomarkers
Performance of the AI model in predicting selected molecular alterations from H\&E whole-slide images, including but not limited to IDH1/2 mutation, 1p/19q codeletion, H3 K27M/G34 alteration, TERT promoter mutation, and BRAF V600E, measured by AUC, sensitivity, specificity, and MCC.
Time frame: Up to Jul 2029
Agreement between AI attention maps and neuropathologist-identified diagnostic regions
Qualitative and semi-quantitative interpretability assessment of overlap between model attention heatmaps and diagnostically relevant regions identified independently by expert neuropathologists.
Time frame: Up to Jul 2029
Human versus AI versus AI-assisted diagnostic performance
Comparison of diagnostic accuracy, inter-rater agreement, and slide review time among AI-only diagnosis, pathologist-only diagnosis, and AI-assisted pathologist diagnosis on a selected internal test subset. Inter-rater agreement will be evaluated using Cohen's kappa where appropriate.
Time frame: Up to Jul 2029
This platform is for informational purposes only and does not constitute medical advice. Always consult a qualified healthcare professional.