Computational Complexity Reduction of Neural Networks of Brain Tumor Image Segmentation by Introducing Fermi-Dirac...

Area of Science:

• Computational neuroscience
• Medical image analysis
• Artificial intelligence

Background:

• Deep learning models require substantial computational resources, often necessitating high-cost hardware.
• Advancements in deep learning are hindered by the reliance on expensive computational infrastructure.
• Developing efficient preprocessing methods is crucial for democratizing deep learning.

Purpose of the Study:

• To introduce a novel preprocessing method inspired by solid-state physics to reduce neural network computational complexity.
• To develop a Fermi-Dirac correction function for voxel intensity normalization and filtering.
• To validate the method's performance and computational efficiency on medical imaging datasets.

Main Methods:

• Image space is isomorphically mapped to a non-interacting physical system, treating voxels as particle-like clusters.
• A Fermi-Dirac distribution is reconstructed as a correction function for voxel intensity normalization and filtering.
• The method was validated using the BraTS 2019 dataset and a dimensional fusion U-net architecture.

Main Results:

• The Fermi-Dirac correction function demonstrated comparable performance to existing preprocessing methods.
• The proposed algorithm achieved at least a 38% reduction in computational time compared to z-score and Gamma correction on low-cost hardware.
• It offers superior image augmentation and segmentation capabilities compared to global histogram equalization.

Conclusions:

• The Fermi-Dirac correction function effectively reduces computational complexity in deep learning models.
• This approach offers a viable alternative to high-cost hardware, making deep learning more accessible.
• The method shows promise for enhancing image analysis tasks in medical imaging and beyond.