Data-dependent Nonlinearity Analysis in CT Denoising CNNs.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning methods are increasingly used in medical imaging for formation, processing, and analysis.
• The nonlinear nature of neural networks presents challenges in analyzing image properties, generalizability, and robustness.
• Conventional methods struggle to capture the complex data-dependent characteristics of neural network outputs.

Purpose of the Study:

• To analyze a class of neural networks utilizing piece-wise linear activation functions.
• To develop a metric for quantifying the fidelity of local linear approximations in trained neural network models.
• To enable analytical prediction and quantitation of performance for specific input data.

Main Methods:

• Representing neural networks with piece-wise linear activation functions as locally linear systems.
• Developing and applying a nonlinearity index metric to assess local linear approximation fidelity.
• Analyzing three computed tomography (CT) denoising convolutional neural networks (CNNs) to predict output noise properties.

Main Results:

• The proposed nonlinearity index metric shows a high correlation with the accuracy of noise predictions in output images.
• The analysis successfully predicted noise properties in CT denoising CNNs.
• The locally linear system representation allows for standard propagation methods to estimate image noise.

Conclusions:

• The developed nonlinearity index provides theoretical understanding of neural network behavior in medical imaging.
• This approach enables performance prediction and quantitation for specific data inputs.
• The findings contribute to improving the reliability and interpretability of deep learning in medical image analysis.