Federated Gastrointestinal Lesion Classification with Clinical-Entropy Guided Quantum-Inspired Token Pruning in...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computational Biology

Background:

• Gastrointestinal (GI) cancers are a significant global health issue.
• Accurate endoscopic interpretation is crucial for patient outcomes.
• Deep learning aids GI diagnosis but faces privacy and data challenges.

Purpose of the Study:

• To develop a privacy-preserving federated learning framework for GI endoscopy analysis.
• To enhance the efficiency of Vision Transformers (ViTs) in federated settings.
• To address computational and communication costs in federated ViT models.

Main Methods:

• Proposed a federated framework combining ViTs with a Clinical-Entropy Guided Quantum Evolutionary Algorithm (CEQEA).
• CEQEA enables adaptive token pruning based on local data diversity.
• Framework trained on HyperKVASIR dataset under non-IID conditions.

Main Results:

• Achieved 92.33% accuracy for lower-GI and 90.19% for upper-GI classification.
• Maintained high specificity across diagnostic classes.
• Adaptive pruning reduced token processing by ~40% and communication rounds by 33%.

Conclusions:

• Entropy-aware, quantum-inspired evolutionary optimization balances performance and efficiency.
• Transformer-based models are practical for privacy-preserving, multi-institutional GI endoscopy.
• The framework enhances AI adoption in clinical settings while preserving patient privacy.