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

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Oncology research

Background:

• Accurate glioma cell classification is hindered by poor feature representation, limited morphological analysis, and image artifacts.
• Existing methods struggle with fine-grained pathological details and robustness against staining variations and noise.

Purpose of the Study:

• To develop a hybrid classification framework for enhanced glioma cell recognition.
• To improve fine-grained feature representation and morphological relationship modeling.
• To increase classification robustness against imaging interferences.

Main Methods:

• A hybrid framework combining a convolutional residual local detail enhancement front-end and a Transformer backbone.
• Integration of a morphological relation modeling mechanism and a causal-invariant discrimination module.
• Enhancement of nuclear boundaries, staining textures, and local heterogeneity while suppressing noise.

Main Results:

• Achieved 98.46% accuracy, 97.94% precision, and 98.07% recall on a dataset of 473 subjects.
• Demonstrated superior performance compared to competing methods.
• Ablation and visualization studies confirmed the effectiveness of individual modules.

Conclusions:

• The proposed method significantly improves fine-grained feature representation and structural relation modeling for glioma cells.
• It offers enhanced classification robustness, providing reliable technical support for intelligent pathological cell recognition.
• This approach advances the field of computational pathology for brain tumor diagnosis.