© Wuhan University 2026

0 Introduction

Precise and efficient medical image segmentation is paramount in medical image analysis ^[1-2]. However, conventional manual segmentation is labor-intensive, time-consuming, and prone to inter-expert variability ^[3-4]. This necessitates automated approaches, particularly deep learning algorithms, to accurately delineate organs or pathological regions. Such methods are crucial for facilitating accurate, rapid, and consistent diagnoses for clinicians and researchers^[5]. In recent years, the proliferation of advanced deep learning architectures, including Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and Mamba, has led to substantial breakthroughs in medical image segmentation. Nonetheless, despite their notable benefits, each architecture faces inherent performance constraints dictated by its intrinsic design attributes ^[6]. CNNs, relying on local convolutional kernels, inherently struggle to capture long-range dependencies, which can lead to suboptimal feature extraction and segmentation outcomes. Conversely, ViTs, while adept at global modeling, are hampered by quadratic computational complexity, limiting their efficiency in dense prediction tasks. Furthermore, Mamba-based models encounter challenges in achieving an optimal effective receptive field ^[7] when converting 2D image data into 1D sequential formats.

The Receptance Weighted Key Value (RWKV) model^[8] introduces the WKV attention mechanism alongside token shift layers, facilitating linear computational complexity within global attention mechanisms while effectively capturing local dependencies. Despite the innovative efforts to extend the application of RWKV to visual domains, challenges persist in the direct adaptation of RWKV for dense image prediction tasks, including the nuanced field of medical image segmentation. This primarily stems from the inherent incompatibility between the causal sequential modeling capabilities of the RWKV architecture and the 2D spatial structure of images. Tailored for 1D sequences modeling, RWKV is not inherently suited for direct application to model 2D image tokens. Prior research has addressed this issue by flattening 2D tokens into 1D sequences via a computational hierarchy^[9], such as row-column-major ordering, where each row's end is immediately followed by the next row's start. However, this approach ignores the preservation of spatial continuity^[10], thereby compromising the integrity of the intrinsic structural information of the image.

This paper introduces an optimized medical image segmentation framework, termed Multi-head Scan RWKV (MS-RWKV), which is an adaptation of the RWKV architecture ^[8] for 2D visual tasks. The proposed model preserves the fundamental architecture and inherent benefits of RWKV, while incorporating essential modifications tailored for the segmentation of 2D medical images. 1) Our multi-head scanning module reduces the unidirectional causal bias among image patches, enabling more balanced global receptive field computation. We strategically insert padding tokens between scan-sequence elements that are spatially disconnected, preserving 2D structural continuity. 2) Building on the GhostNetV2 ^[11] architecture, we design a feature aggregation block that captures local spatial context while strengthening correlations within 1D sequences generated along specific scan directions, using asymmetric convolutions. 3) This novel mechanism broadens token semantics by aggregating multi-scale features from wide receptive fields, effectively addressing orientation sensitivity in 2D images. During training, panoramic token shift (P-Shift) employs structural reparameterization to learn adaptive token shifting across diverse contexts.

We rigorously evaluated our proposed MS-RWKV model through comprehensive experiments on skin lesion segmentation and multi-organ segmentation tasks, showcasing its superior performance and high efficiency in the field of medical image segmentation. Ad-ditionally, we conducted ablation studies to validate the performance of our design. An extensive array of experimental outcomes underscores the robust potential of our model for image segmentation applications.

The main contributions of this study are as follows:

• We introduced the MS-RWKV framework, adapting the RWKV model for application in medical image segmentation tasks. This adaptation has demonstrated promise as an enhanced solution for more precise and effective image segmentation.

• We integrated a novel multi-head scan mechanism, augmented with padding strategies, into the RWKV architecture. This innovation effectively bridges the divide between 1D sequences processing and 2D image traversal.

• During the conversion of 2D images to 1D sequences, we integrate the Feature Aggregation Attention (FAA) module. The asymmetric convolution within this module extracts features that are particularly advantageous for subsequent processing of 1D sequences.

1 Methodology

The RWKV model^[12], drawing from natural language processing, combines the parallel training efficiency of transformers^[13] with the sequential inference capabilities of Recurrent Neural Networks (RNNs)^[14]. The architecture of the RWKV model comprises a series of stacked blocks, each of which integrates time-mixing and channel-mixing blocks, both of which feature recurrent structures.

Time-Mixing Block. This block is engineered to augment the modeling capacity for dependencies and patterns within sequential data. Given an input sequence $\mathit{x}=(x_{\mathrm{1}},x_{\mathrm{2}},\dots ,x_T)$, where T represents the length of the input features after convolutional subsampling, the output sequence $\mathit{o}=(o_{\mathrm{1}},o_{\mathrm{2}},\dots ,o_T)$ of the time-mixing block is computed as follows:

${\mathit{r}}_t=({\mu }_{\mathrm{r}}\odot x_t+(\mathrm{1}-{\mu }_{\mathrm{r}})\odot x_{t-\mathrm{1}})\cdot {\mathit{W}}_{\mathrm{r}},$(1)

${\mathit{k}}_t=({\mu }_{\mathrm{k}}\odot x_t+(\mathrm{1}-{\mu }_{\mathrm{k}})\odot x_{t-\mathrm{1}})\cdot {\mathit{W}}_{\mathrm{k}},$(2)

${\mathit{v}}_t=({\mu }_{\mathrm{v}}\odot x_t+(\mathrm{1}-{\mu }_{\mathrm{v}})\odot x_{t-\mathrm{1}})\cdot {\mathit{W}}_{\mathrm{v}},$(3)

${\mathit{o}}_t=(\sigma (r_t)\odot \mathrm{w}\mathrm{k}{\mathrm{v}}_t)\cdot {\mathit{W}}_{\mathrm{o}},$(4)

where ${\mathit{W}}_{\mathrm{o}}\in {\mathbb{R}}^{d_{\mathrm{i}\mathrm{o}}\times d_{\mathrm{a}\mathrm{t}\mathrm{t}}}$ is the output projection matrix,$d_{\mathrm{i}\mathrm{o}}$ is the input/output size, and $d_{\mathrm{a}\mathrm{t}\mathrm{t}}$ is the RWKV time-mixing block size. ${\mathit{W}}_{\mathrm{r}}\in {\mathbb{R}}^{d_{\mathrm{a}\mathrm{t}\mathrm{t}}\times d_{\mathrm{i}\mathrm{o}}}$, ${\mathit{W}}_{\mathrm{k}}\in {\mathbb{R}}^{d_{\mathrm{a}\mathrm{t}\mathrm{t}}\times d_{\mathrm{i}\mathrm{o}}}$ and ${\mathit{W}}_{\mathrm{v}}\in {\mathbb{R}}^{d_{\mathrm{a}\mathrm{t}\mathrm{t}}\times d_{\mathrm{i}\mathrm{o}}}$ are the projection matrices for the receptance, key, and value, respectively. ${\mu }_{\mathrm{r}}$, ${\mu }_{\mathrm{k}}$ and ${\mu }_{\mathrm{v}}$ are time-mixing factors for the receptance, key, and value, respectively. The values of ${\mathit{r}}_t$, ${\mathit{k}}_t$, and ${\mathit{v}}_t$ are calculated through linear interpolation between the current input and the input from the previous time step. This block applies a non-linear activation function σ to the receptance vector $r_t$, and then combines the resulting values with the hidden state $\mathrm{w}\mathrm{k}{\mathrm{v}}_t$ through element-wise multiplication.

$\mathrm{w}\mathrm{k}{\mathrm{v}}_t=\frac{{\displaystyle \sum _{i=\mathrm{1}}^{t-\mathrm{1}}}{\mathrm{e}}^{-(t-\mathrm{1}-i)\mathit{w}+k_i}{v}_i+{\mathrm{e}}^{u+k_t}{v}_t}{{\displaystyle \sum _{i=\mathrm{1}}^{t-\mathrm{1}}}{\mathrm{e}}^{-(t-\mathrm{1}-i)\mathit{w}+k_i}+{\mathrm{e}}^{u+k_t}},$(5)

where $\mathit{w}$ is the channel-wise time decay vector for the previous input, $u$ is the special weighting factor applied to the current input, and $\mathrm{w}\mathrm{k}{\mathrm{v}}_t$ is the weighted summation of the input in the interval [1, t]. The hidden states Eq. (5) can be computed recursively as follows:

$\mathrm{w}\mathrm{k}{\mathrm{v}}_t=\frac{a_{t-\mathrm{1}}+{\mathrm{e}}^{u+k_t}{v}_t}{b_{t-\mathrm{1}}+{\mathrm{e}}^{u+k_t}},$(6)

where $a_t={\mathrm{e}}^{-w}{a}_{t-\mathrm{1}}+{\mathrm{e}}^{k_t}{v}_t,b_t={\mathrm{e}}^{-w}{b}_{t-\mathrm{1}}+{\mathrm{e}}^{k_t},$ and $a_{\mathrm{0}},b_{\mathrm{0}}$ are zero-initialized.

Channel-Mixing Block. This block is specifically engineered to enhance the feature representations propagated from the time-mixing block through a series of robust non-linear transformations. Given the input sequence ${\mathit{x}}^{\text{'}}=(x_{\mathrm{1}}^{\mathrm{\text{'}}},x_{\mathrm{2}}^{\mathrm{\text{'}}},\dots ,x_T^{\mathrm{\text{'}}})$, the specific process of the block is:

$r_t^{\mathrm{\text{'}}}=({\mu }_r^{\mathrm{\text{'}}}\odot x_t^{\mathrm{\text{'}}}+(\mathrm{1}-{\mu }_r^{\mathrm{\text{'}}})\odot x_{t-\mathrm{1}}^{\mathrm{\text{'}}})\cdot {\mathit{W}}_{\mathrm{r}}^{\mathrm{\text{'}}},$(7)

${\mathit{k}}_t^{\mathrm{\text{'}}}=({\mu }_k^{\mathrm{\text{'}}}\odot x_t^{\mathrm{\text{'}}}+(\mathrm{1}-{\mu }_k^{\mathrm{\text{'}}})\odot x_{t-\mathrm{1}}^{\mathrm{\text{'}}})\cdot {\mathit{W}}_{\mathrm{k}}^{\mathrm{\text{'}}},$(8)

${\mathit{o}}_t^{\mathrm{\text{'}}}=\sigma (r_t^{\mathrm{\text{'}}})\odot (\mathrm{m}\mathrm{a}\mathrm{x}(k_t^{\mathrm{\text{'}}}{,\mathrm{0})}^{\mathrm{2}}\cdot {\mathit{W}}_{\mathrm{v}}^{\mathrm{\text{'}}}),$(9)

where ${\mathit{W}}_{\mathrm{r}}^{\mathrm{\text{'}}}\in {\mathbb{R}}^{d_{\mathrm{l}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{a}\mathrm{r}}\times d_{\mathrm{i}\mathrm{o}}}$ and ${\mathit{W}}_{\mathrm{k}}^{\mathrm{\text{'}}}\in {\mathbb{R}}^{d_{\mathrm{l}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{a}\mathrm{r}}\times d_{\mathrm{i}\mathrm{o}}}$ are the projection matrices for the receptance and key, respectively. ${\mathit{W}}_{\mathrm{v}}^{\mathrm{\text{'}}}\in {\mathbb{R}}^{d_{\mathrm{l}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{a}\mathrm{r}}\times d_{\mathrm{i}\mathrm{o}}}$ is the channel-mixing matrix, and $d_{\mathrm{l}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{a}\mathrm{r}}$ is the RWKV time-mixing block size. ${\mu }_r^{\mathrm{\text{'}}}$ and ${\mu }_k^{\mathrm{\text{'}}}$ are time-mixing factors for the receptance and key, respectively. The channel-mixing block operates causally, as the computation of o_t^′ is contingent solely on x_t^′ and x_t-1^′ . Intuitively, this amplification process enhances the representations of historical information.

2 Architecture of MS-RWKV

The RWKV^{[15, 12]} architecture, originally conceived for processing 1D sequences, encoun-ters some limitations when tasked with learning from 2D data structures. To address these challenges, we introduce novel modules that enhance the RWKV's capability to effectively process 2D image data.

Overall architecture. The architecture of the MS-RWKV is depicted in Fig. 1(a). The MS-RWKV incorporates a four-stage hierarchical backbone with skip connections. Given an input image I, we first partition the feature map $X\in {\mathbb{R}}^{H\times W\times \mathrm{3}}$ into 2D patches via the non-overlapping patch embedding layer, with the channel dimension projected into c dimensions. As illustrated in Fig.1 (b), each stage's MS-RWKV module is tasked with subsequently extracting feature representations at varying levels from the input image. Within these modules, four distinct multi-head scanning trajectories are utilized to linearize the patch tokens into sequences X= [S₁, S₂, S₃, S₄], where S_n is the sequence after the n-th scan path. The MS-RWKV module achieves competitive performance by computing global and local attention with linear complexity for input sequences. Following the decoding phase, the image resolution is restored to its original size through the final projection layer, enabling pixel-accurate segmentation. A comprehensive exposition of our architectural design principles and methodological implementations will be systematically delineated in the following sections.

Fig. 1 (a) The overall architecture of MS-RWKV. (b) The core of MS-RWKV block

Feature Aggregation Module. The RWKV^[12] model, initially tailored for 1D input sequences, encounters challenges in preserving local dependency relationships when applied to 2D image data, which impedes its ability to capture local fine-grained details^[16]. Building upon GhostNetV2's^[11] approach to capturing local details, we develop a novel feature aggregation module, Feature Aggregation Attention (FAA), to enhance feature aggregation in dense prediction tasks. This module is engineered to enhance local feature extraction capability by expanding effective receptive fields while preserving parameter sparsity through grouped convolutions. FAA consists of two Multi-scale Convolutional Modules (MCM) and an Activation Module (AM). It can be described as follows:

${\widehat{x}}_i=\mathrm{M}\mathrm{C}\mathrm{M}(\mathrm{D}\mathrm{W}\mathrm{C}(\mathrm{M}\mathrm{C}\mathrm{M}(x_i)\odot \mathrm{A}\mathrm{M}(x_i)))+x_i,$(10)

where x_i and ${\widehat{x}}_i$ represent the input and output tensors of the module in the i-th stage, DWC is a depth-wise convolution with a kernel size of 3 × 3.

As illustrated in Fig. 2 , the input x is partitioned into four subcomponents x = (X₁, X₂, X₃, X₄), which are processed through parallel convolutional pathways with asymmetric kernel dimensions. The resultant features are concatenated to form the final output, achieving multi-granularity feature fusion that captures both local details and global context. Crucially, the asymmetric convolutions (kernel size: 1 × K_H , K_W × 1) explicitly model directional spatial correlations, thereby significantly facilitating sequential scanning in the subsequent RWKV module. The AM branch implements a module consisting of two linear layers and an activation function. FAA leverages the features expanded by the first MCM module, which are then modulated by the AM branch, augmenting the model's expressive power. The enhanced features are subsequently fed into the second MCM module to restore the original feature information for output, effectively aggregating surrounding information. This method effectively mitigates the inherent limitations of the flattening approach.

Fig. 2 (a) The diagrams of blocks in feature aggregation module. (b) Multi-scale convolutional module

Scan patterns. The multi-head scan mechanism, which involves the parallel extraction and integration of features across various divergent scanning trajectories ^[17-18], enables the capture of a global receptive field and the modeling of long-range dependencies. This module also draws inspiration from the research on multi-head scan presented in UltraLight VM-UNet^[19] and MHS-VM^[20], and we conducted correlation experiments to explore the implications and applications of these methods. The details of the aforementioned process can be formulated as follows:

$S_{\mathrm{1}},S_{\mathrm{2}},S_{\mathrm{3}},S_{\mathrm{4}}=\mathrm{S}\mathrm{p}\left[\mathrm{L}\mathrm{N}\left(X_{\mathrm{i}\mathrm{n}}\right)\right],$(11)

$\mathrm{R}{\mathrm{W}}_{-}{S}_i=\mathrm{R}\mathrm{W}\mathrm{K}{\mathrm{V}}_p(\mathrm{P}\mathrm{r}\mathrm{o}\mathrm{j}\left(S_i\right)),i=\mathrm{1,2},\mathrm{3,4},$(12)

$X_o=\mathrm{C}\mathrm{a}\mathrm{t}\left(\mathrm{R}{\mathrm{W}}_{-}{S}_{\mathrm{1}},\mathrm{R}{\mathrm{W}}_{-}{S}_{\mathrm{2}},\mathrm{R}{\mathrm{W}}_{-}{S}_{\mathrm{3}},\mathrm{R}{\mathrm{W}}_{-}{S}_{\mathrm{4}}\right),$(13)

$\mathrm{O}\mathrm{u}\mathrm{t}=\mathrm{P}\mathrm{r}\mathrm{o}\mathrm{j}\left[\mathrm{L}\mathrm{N}\left(X_o\right)\right],$(14)

where LN is the LayerNorm, Sp is the Split operation, RWKV_p is the RWKV operation with padding, Cat is the concatenation operation, and Proj is the Projection operation. The information gathered from the four branches is then merged and cycled through the subsequent model module. Within each layer, consecutive modules integrate various scanning approaches, thereby enhancing the model's generalization capabilities^[21]. Ultimately, we opted for a parallel 4-head scan approach to uniformly decompose the depth feature X_in into four sequences [S₁, S₂, S₃, S₄]. In Eqs.(12) and (13), each branch RWKV_p with padding p independently extracts pertinent information, which is subsequently aggregated and output through a concatenation strategy. To address the spatial discontinuity resulting from the flattening of the image into 1D sequences, we insert padding tokens at the end of each row and column to detect the end-of-line signal. We execute both row-major and column-major scans in both forward and reverse directions, as delineated in Fig.3. To preserve the resolution of the original image, we remove padding tokens prior to concatenation, thereby restoring the image to its input size. Finally, the four features are concatenated to form the composite feature X_o, which is subsequently processed by LayerNorm and Projection operations to yield the Out.

Fig. 3 Illustrations of the scan head block Note:   After padding the image, one of the four scanning methods is selected to flatten the image into 1D sequences.

Token shift. The token shift mechanism in RWKV ^[8] was initially introduced to mitigate the misalignment between the 1D decay of attention and the 2D adjacency in images. However, current token shift methods, including the unidirectional token shift (Uni-shift) in RWKV ^[8] and the quadridirectional token shift (Quad-shift) in Vision-RWKV^[22], collect information from limited directions and fail to account for the comprehensive spatial continuity inherent in 2D imagery. To address this, we employ a panoramic token shift (P-Shift) mechanism, integrating it into time-mixing and channel-mixing blocks in the RWKV module. In our approach, we harness a suite of deep convolutional layers equipped with convolutional kernels of diverse sizes to facilitate the extraction and fusion of features. This strategy effectively aggregates information from multiple spatial orientations, enhancing the model's capacity to capture richer spatial features^[23]. The mechanism of P-Shift can be formulated as follows:

${\widehat{z}}_i=\sum _{\mathrm{k}\mathrm{s}\in \mathrm{K}\mathrm{S}}\mathrm{D}\mathrm{W}\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{\mathrm{k}\mathrm{s}}(z_i)+z_i,$(15)

where $z_i$ and ${\widehat{z}}_i$ represent the input and output tensors of the module in the i-th tage. DWConv_ks denotes a depthwise convolution with a kernel size of ks. KS defines a set of parallel convolution kernels with values of {1 × 1, 3 × 3, 5 × 5}. This mechanism utilizes a multi-branch architecture during training to capture local contextual information with an expanded visual receptive field. For testing, the multi-branch structure ^[24] is consolidated into a single branch using a 5×5 convolution kernel, thereby enhancing the model's inference efficiency and reducing parameter count.

3 Experiments and Results

3.1 Datasets and Parameter

We conducted an extensive performance evaluation of our framework across three distinct open-source medical image segmentation datasets: ISIC17^[25], ISIC18 ^[26] and ACDC^[27]. These benchmark datasets are widely recognized for their pivotal role in advancing medical image segmentation research. In our experimental setup, we implemented the MS-RWKV model using PyTorch 2.0 and performed all training on an NVIDIA GeForce RTX 3090Ti GPU. To enhance the robustness of our model, we incorporated data augmentation techniques, including random flipping and rotation. The training was conducted with a batch size of 32 over 300 epochs. We opted for the AdamW optimizer for network training, complemented by a cosine annealing schedule for learning rate decay. Given the heterogeneity in difficulty across different datasets, we fine-tuned the hyperparameters accordingly. For the ACDC dataset, we initialized the learning rate at 5 × 10^-4 and set the weight decay to 1 × 10^-4. For the remaining datasets, we standardized the learning rates at 1 × 10^-3 and weight decays at 1 × 10^-5.To ensure a rigorous and comprehensive performance evaluation, we adopted a standardized set of quantitative metrics, including the Mean Intersection over Union (mIoU), the Dice Similarity Coefficient (DSC), and Accuracy (Acc).

3.2 Main Results

To verify the effectiveness of our proposed method, we performed a comparative analysis of the MS-RWKV model against several state-of-the-art models.

Table 1 presents our method's performance compared with various approaches on ISIC17 and ISIC18 datasets, where our proposed MS-RWKV achieved the best average mIoU of 83.27% and 81.52%. Specifically, compared with Mamba-based methods (such as H-vmunet^[28]), our method improved the mIoU by 1.22% and 0.92%, respectively, and by 0.56% and 0.47% compared with RWKV-based methods (such as RWKV-UNet^[29]). As shown in Table 2, compared with other methods, our proposed method achieved the best average DSC of 91.85% on the ACDC dataset.

The quantitative results presented in the tables demonstrate that our method achieves superior performance compared to state-of-the-art approaches for 2D medical image segmentation tasks.

3.3 Ablation Studies

We conducted comprehensive ablation studies on the ISIC18 dataset to validate the effectiveness of the multi-head scan, feature aggregation attention, and P-shift components. The results are shown below.

Multi-Head Scan with Padding. In our ablation studies, we adopt four parallel scan heads for hierarchical feature extraction from 2D image data. To systematically evaluate the impact of architectural components, we conducted comparative experiments with two distinct configurations: 1) four scan heads without padding mechanisms, and 2) three scan heads with strategic padding implementation. The quantitative results, as detailed in Table 3, demonstrate that increasing the number of scan heads leads to enhanced model capacity and improved feature representation.

Token Shift. To explore the effectiveness of the proposed P-Shift, we compared its performance with Uni-Shift in RWKV^[8] and Quad-Shift in Vision-RWKV^[22]. The ablation studies presented in Table 4 demonstrate that our proposed P-Shift mechanism coupled with reparameterization significantly enhances the local feature extraction capability of the token shift operation. Our novel token shift architecture effectively exploits the inherent spatial correlations within 2D visual feature maps, facilitating multi-directional feature propagation and adaptive feature aggregation across diverse spatial orientations.

Feature Aggregation Attention. To systematically investigate the impact of various architectural components preceding the MS-RWKV Block, we conducted a series of ablation experiments. Our assessment focused on comparing the performance of two distinct baseline modules (FFN and GhostNetV2) against our proposed FAA module. As presented in Table 5, the proposed FAA module achieves superior performance, surpassing FFN by 1.32% and GhostNetV2 by 0.54% in mIoU.

We conducted comprehensive ablation studies on ISIC18 to evaluate individual components. Table 6 summarizes ablation results, where the last row shows the full model's performance and the preceding rows quantify the impact of module removal. A comparative analysis between the full model and the configurations lacking the FAA reveals that FAA contributes most significantly to the overall performance.

3.4 Visualization

Qualitative analysis of ISIC 2018 results (Fig. 4) demonstrates that our MS-RWKV architecture facilitates the feature integration of semantic information and finer-grained features. Quantitative experiments confirm that our proposed framework achieves significant gains over the VM-UNet baseline, particularly in capturing subtle texture variations as confirmed by quantitative metrics. These visualizations further validate that MS-RWKV, as an RWKV-based model, holds significant potential in the field of medical image segmentation.

Fig. 4 The visual comparison of segmentation results of ours model and other segmentation methods against ground truth on ISIC2018 dataset

4 Conclusion

In this paper, we propose MS-RWKV, which extends the foundational architecture of the RWKV model. MS-RWKV enhances RWKV through three innovations: multi-head scanning with adaptive padding for 2D coherence, P-Shift for spatial dependencies, and feature aggregation attention for multi-scale fusion. Comprehensive empirical evaluations across diverse medical image segmentation benchmarks demonstrate that our MS-RWKV architecture achieves superior performance compared to state-of-the-art approaches. We further plan to extend MS-RWKV to tasks beyond segmentation, such as cross-modal registration and high-fidelity reconstruction.

References

All Tables

All Figures

	Fig. 1 (a) The overall architecture of MS-RWKV. (b) The core of MS-RWKV block
In the text

	Fig. 2 (a) The diagrams of blocks in feature aggregation module. (b) Multi-scale convolutional module
In the text

	Fig. 3 Illustrations of the scan head block Note:   After padding the image, one of the four scanning methods is selected to flatten the image into 1D sequences.
In the text

	Fig. 4 The visual comparison of segmentation results of ours model and other segmentation methods against ground truth on ISIC2018 dataset
In the text