The paper introduces Mixture-of-Wavelet-Experts (MoWE), a pooling approach designed for U-Net architectures in ultrasound lesion segmentation. MoWE uses spatially adaptive routing to select among wavelet-based experts, rather than applying a fixed downsampling operation everywhere.
Standard pooling operations suppress information, according to the abstract, which motivates the shift to a more adaptive mechanism. MoWE is proposed as a direct replacement for max-pooling and strided convolutions, though the available text does not include experimental results or comparisons.
This digest is based on a single arXiv abstract; no independent findings or contrasting perspectives are reported.