Quadratic Autoencoder (Q-AE) for Low-Dose CT Denoising.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Medical Imaging

Background:

• Current artificial neurons use inner products, limiting individual neuron capability.
• Biological neurons exhibit complexity and diversity, inspiring novel artificial neuron designs.
• Prior theoretical work highlights quadratic neurons' merits in representation, efficiency, and interpretability.

Purpose of the Study:

• To evaluate the efficacy of quadratic neurons in deep learning architectures.
• To introduce and apply a quadratic autoencoder for low-dose computed tomography (CT) denoising.
• To demonstrate the potential of quadratic-neuron-based deep learning in medical imaging.

Main Methods:

• Developed a novel encoder-decoder structure using quadratic neurons, termed the quadratic autoencoder.
• Applied the quadratic autoencoder to the task of low-dose CT image denoising.
• Utilized the Mayo low-dose CT dataset for experimental validation.

Main Results:

• The quadratic autoencoder demonstrated significant utility and robustness in low-dose CT image denoising.
• The model achieved high efficiency in terms of performance and computational resources.
• Experimental results validate the effectiveness of quadratic neurons in a practical medical imaging application.

Conclusions:

• Quadratic neurons offer enhanced capabilities compared to traditional artificial neurons.
• The quadratic autoencoder is a promising new tool for medical image denoising, particularly for low-dose CT.
• This work represents a novel implementation of deep learning with new neuron types, showing significant potential in medical imaging.