Multimedia Systems Computer Networks and Communications
GaborMix: Adaptive Gabor configuration mixing for boundary-aware medical image segmentation
Abstract
Medical image segmentation on 2D images remains challenging, largely due to the ambiguous boundaries between target structures and surrounding tissues. To address this issue, we enhance the boundary representation capability of segmentation models by revisiting Gabor filters as explicit boundary encoders. Specifically, we introduce GaborMix, a search-free Gabor configuration mixer for 2D segmentation that generates multiple oriented responses from a small set of kernel-size and orientation anchors ( k , d ), and employs lightweight, feature-conditioned gating to adaptively mix predefined spatial-support and orientation-resolution anchors, while spatially varying directional aggregation enhances features around indistinct boundaries. We demonstrate the versatility of GaborMix by integrating it into three popular backbones: U-Net, Res2Net, and PVTv2, and observing consistent performance gains. Building on the strong results with PVTv2, we further develop GaborMix-PVT, a robust network for a broad range of medical image segmentation tasks. Extensive experiments on eight 2D benchmarks spanning four categories (polyp, skin lesion, breast ultrasound, and cytopathology) show that our method achieves the best average performance while remaining competitive on individual datasets.