PlexusNet

PlexusNet implements a scalable neural architecture for clinical image classification that optimizes network depth, width, and branching via neural architecture search and employs learnable data normalization to reduce parameters and computational cost while maintaining competitive accuracy.

Key Features:

• Scalable Architecture: The framework adjusts depth, width, and branching structure to optimize performance across clinical classification problems while maintaining computational efficiency.
• Efficient Model Design: The architecture reduces parameters and feature maps compared to ResNet-18 and EfficientNet B0/1, achieving competitive accuracy with lower computational cost.
• Learnable Data Normalization: A data normalization algorithm dynamically adapts to dataset characteristics during training, enhancing generalization without relying on fixed pre-training techniques.
• Neural Architecture Search (NAS): A NAS strategy customizes the architecture for different clinical classification tasks to tailor performance to specific applications.

Scientific Applications:

• Clinical image classification: Applied to five distinct clinical classification problems in medical imaging, demonstrating non-inferior performance to ResNet-18 and EfficientNet B0/1 while using fewer computational resources.

Methodology:

Systematic scaling and design regulate network depth, width, and branching to balance complexity and resource efficiency; visualization of latent features reveals category-specific clusters for interpretability.