A structure-aware and causal-invariant framework for glioma cell classification in pathological images

Accurate classification of glioma cells and normal cells remains challenging due to insufficient fine-grained pathological feature representation, limited modeling of morphological relationships, and interference from staining variation, background noise, and imaging bias.

This study proposes a hybrid classification framework that integrates a convolutional residual local detail enhancement front-end, a Transformer backbone, a morphological relation modeling mechanism, and a causal-invariant discrimination module. The framework enhances nuclear boundaries, staining textures, and local heterogeneity, models cross-regional contextual and morphological correlations, and suppresses non-essential disturbance factors.

Experiments on a dataset of 473 subjects and 859 pathological images show that the proposed method achieves 98.46% accuracy, 97.94% precision, 98.07% recall, and 0.987 AUC, outperforming competing methods. Ablation and visualization results further confirm the effectiveness of the key modules.

The proposed method improves fine-grained feature representation, structural relation modeling, and classification robustness, providing reliable technical support for intelligent glioma-related pathological cell recognition.