Hyper-convolutions via implicit kernels for medical image analysis.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) are foundational for computer vision.
• CNN performance relies on convolutional kernels, with capacity tied to learnable weights.
• Kernel size and channels determine the number of weights in standard CNNs.

Purpose of the Study:

• Introduce hyper-convolutions as a novel CNN building block.
• Decouple kernel size from the number of learnable parameters.
• Enhance CNN flexibility, performance, and robustness.

Main Methods:

• Developed hyper-convolutions that implicitly encode kernels using spatial coordinates.
• Weights in hyper-convolutions are correlated via an encoder mapping coordinates to values.
• Replaced standard convolutional kernels with hyper-convolutions in experimental models.

Main Results:

• Hyper-convolutions improved performance compared to standard convolutions.
• Achieved better results with a reduced number of learnable parameters.
• Demonstrated increased robustness against noise in experimental evaluations.

Conclusions:

• Hyper-convolutions offer a flexible and efficient alternative to standard convolutional kernels.
• This novel approach enhances CNN performance and robustness.
• The method allows for more adaptable deep learning architectures.