Title: Med-URWKV{\dag}: Advancing Pretrained Pure VRWKV Architectures for Medical Image Segmentation

Medical image segmentation serves as a cornerstone for computer-aided diagnosis and therapeutic planning. Despite the prevalence of current methodologies leveraging CNNs, ViTs, Mamba, and hybrid structures, these approaches are often constrained by issues including limited receptive fields, excessive computational demands, or suboptimal accuracy. Recently, Vision Receptive-field Weighted Key-Value (VRWKV) models have gained traction as a viable alternative, offering robust capabilities for modeling long-range dependencies in visual tasks. Nevertheless, existing research on VRWKV applications in medical image segmentation has predominantly centered on hybrid architectures trained from scratch, leaving the potential of large-scale, pretrained pure VRWKV models largely unexamined.

This study conducts a systematic evaluation of pure VRWKV architectures for medical image segmentation. We introduce Med-URWKV-T and Med-URWKV-S, which are built by combining pretrained VRWKV encoders of varying scales with pure VRWKV decoders, thereby facilitating a thorough assessment of pretrained pure VRWKV models within this specific domain. To boost performance further, we introduce two modules compatible with VRWKV: the Frequency-Aware Wavelet Attention (FAWA) module, which utilizes wavelet transforms to extract edge details and structural features, and the Multi-Scale Channel Fusion (MSCF) module, which merges multi-scale features to reinforce informative channel representations. The integration of these components into Med-URWKV-T results in the enhanced Med-URWKV{\dag}.

Comprehensive experiments across five medical image segmentation datasets reveal that Med-URWKV delivers performance on par with or exceeding state-of-the-art techniques and meticulously crafted hybrid VRWKV designs. Furthermore, Med-URWKV{\dag} elevates segmentation accuracy beyond that of Med-URWKV-S, accomplishing this with merely half the number of parameters, and secures the highest average Dice similarity coefficient of 88.00%. The associated code will be made publicly available.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC