Related Experiment Video
Updated: Sep 12, 2025
05:47
Evidence-based Knowledge Synthesis and Hypothesis Validation: Navigating Biomedical Knowledge Bases via Explainable AI and Agentic Systems
Published on: June 13, 2025
578
Distilling knowledge from graph neural networks trained on cell graphs to non-neural student models
Vasundhara Acharya1, Bülent Yener2, Gillian Beamer3
1Rensselaer Polytechnic Institute, Troy, USA. acharv2@rpi.edu.
Scientific Reports
|August 10, 2025
Summary
This study introduces biologically informed cell graphs and knowledge distillation for pathology AI. It shows that even non-neural models can learn from AI teachers, improving diagnostic feature importance.
Area of Science:
- Computational pathology
- Artificial intelligence in medicine
- Graph Neural Networks
Background:
- Whole Slide Imaging (WSI) analysis is limited by patch-wise approaches that miss cell interactions.
- Cell graph representations offer a more detailed analysis of local cell interactions, crucial for prognostic value.
- Graph Neural Networks (GNNs) show promise in analyzing these spatial features for classification tasks.
Purpose of the Study:
- To develop improved cell graph construction methods using biologically informed criteria.
- To explore knowledge distillation (KD) where non-neural student models learn from neural network teachers.
- To evaluate the effectiveness of this approach across diverse dataset complexities and identify key diagnostic features.
Main Methods:
- Constructed cell graphs with biologically informed edge thresholds, moving beyond simple density/sparsity.
- Implemented knowledge distillation where a neural network teacher guides non-neural student models.
- Evaluated performance on datasets with varying complexities, including distribution shifts and class imbalance.
- Compared guidance from softened probabilities (calibrated logits) versus raw logits.
Main Results:
- Biologically informed cell graphs capture meaningful cell interactions.
- Non-neural student models successfully learned from neural network teachers via knowledge distillation.
- Teacher guidance proved effective, especially with distribution shifts, mitigating overfitting.
- Student models emphasized morphological features, aligning with pathologist priorities, particularly in Tuberculosis datasets.
Conclusions:
- The proposed method enhances cell graph analysis and knowledge distillation for computational pathology.
- Non-neural models can effectively learn from complex neural network teachers, broadening AI applications.
- This approach improves feature interpretability and aligns AI-driven insights with clinical expertise.
