Hyper-Convolutions via Implicit Kernels for Medical Imaging

TL;DR

This paper introduces hyper-convolutions, a novel convolutional building block that encodes kernels implicitly via spatial coordinates, allowing for flexible architecture design, improved performance, fewer parameters, and increased noise robustness in medical imaging tasks.

Contribution

It proposes hyper-convolutions that decouple kernel size from learnable parameters, enhancing flexibility and efficiency in CNN architectures for medical imaging.

Findings

Hyper-convolutions improve performance over standard convolutions.

They require fewer parameters for similar or better results.

Hyper-convolutions increase robustness against noise.

Abstract

The convolutional neural network (CNN) is one of the most commonly used architectures for computer vision tasks. The key building block of a CNN is the convolutional kernel that aggregates information from the pixel neighborhood and shares weights across all pixels. A standard CNN's capacity, and thus its performance, is directly related to the number of learnable kernel weights, which is determined by the number of channels and the kernel size (support). In this paper, we present the \textit{hyper-convolution}, a novel building block that implicitly encodes the convolutional kernel using spatial coordinates. Hyper-convolutions decouple kernel size from the total number of learnable parameters, enabling a more flexible architecture design. We demonstrate in our experiments that replacing regular convolutions with hyper-convolutions can improve performance with less parameters, and increase robustness against noise. We provide our code here: \emph{https://github.com/tym002/Hyper-Convolution}