Hyper-Convolutions via Implicit Kernels for Medical Imaging
This addresses a bottleneck in medical imaging by enabling more flexible CNN architectures with reduced computational cost, though it appears incremental as it builds on existing convolution methods.
The paper tackles the limitation of standard CNNs where performance is tied to the number of learnable kernel weights by introducing hyper-convolutions, which implicitly encode kernels using spatial coordinates to decouple kernel size from parameters, resulting in improved performance with fewer parameters and increased noise robustness.
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}