Task-specific neural networks for medical imaging using pretrained fragments

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biomedical Imaging

Background:

• Neural Architecture Search (NAS) traditionally requires extensive computation.
• StitchNet offers a novel approach by assembling networks from pre-trained fragments.
• This method reduces computational cost and enables efficient task-specific model creation.

Purpose of the Study:

• To evaluate the practical application of the StitchNet framework for medical image classification.
• To assess StitchNet's performance in classifying retinal images for Age-Related Macular Degeneration (AMD).
• To validate and enhance StitchNet's capabilities by integrating recurrent neural networks (RNNs) and autoencoders.

Main Methods:

• Utilized the StitchNet framework for neural network construction.
• Employed fragments from five pre-trained networks.
• Integrated recurrent neural networks (RNNs) and autoencoders for enhanced functionality.
• Assessed performance on a dataset of retinal images (healthy, dry AMD, wet AMD).

Main Results:

• StitchNet achieved competitive accuracy with significantly reduced computational overhead.
• Domain-specific optimizations improved StitchNet's adaptability and efficiency.
• The developed StitchNet model outperformed a scientist-designed network by 6% in accuracy.
• Demonstrated the framework's effectiveness for classifying retinal pathologies like drusen and choroidal neovascularization (CNV).

Conclusions:

• StitchNet is a viable and efficient framework for medical image classification tasks.
• The approach offers a promising alternative to traditional NAS, especially for specialized applications like AMD detection.
• Future work will focus on further improving algorithm efficiency and reducing data requirements.