2510.0087 EndoNet: Content-Aware Linear Attention for Endoscopic Video Super-Resolution v1

🎯 ICAIS2025 Accepted Paper
Back to Paper List

Paper Content

📄 Open in New Tab

🎓 Meta Review & Human Decision

Decision:

Accept

Meta Review:

AI Review from DeepReviewer

AI Review available after:
--d --h --m --s

📋 AI Review from DeepReviewer will be automatically processed

📋 Summary

This paper introduces EndoNet, a novel framework for endoscopic video super-resolution (EVSR) that leverages the Receptance Weighted Key Value (RWKV) architecture for efficient long-range temporal modeling and a Dynamic Group-wise Shift (DGW-Shift) mechanism for adaptive spatial kernel composition. The proposed method aims to address the unique challenges of EVSR, such as rapid camera motion, non-rigid tissue deformation, and specular highlights, which are not adequately handled by existing CNN- and transformer-based models. EndoNet integrates these innovations into both temporal and spatial modules, achieving a balance between global context modeling and local adaptability. The authors demonstrate the effectiveness of their approach through extensive experiments on a synthetic endoscopic video dataset, showing that EndoNet achieves consistently strong performance over recent baselines in quantitative comparisons. However, the paper's evaluation is limited to a single dataset, and there is a lack of detailed analysis of computational complexity and efficiency, which raises concerns about the practical applicability of the method in real-world clinical settings. The paper also lacks a thorough discussion of the limitations of the proposed method, including potential failure cases and scenarios where it may not perform well. Despite these limitations, the paper makes a significant contribution to the field of medical video enhancement by introducing the RWKV architecture and the DGW-Shift mechanism, which are novel and promising approaches for EVSR.

✅ Strengths

One of the most compelling aspects of this paper is the introduction of the RWKV architecture to the EVSR field, which is a novel application in the domain of medical video enhancement. The RWKV architecture, a linear-complexity, transformer-RNN hybrid, is particularly well-suited for modeling long-range temporal dependencies, which are crucial in endoscopic videos due to the rapid and complex motion patterns. The authors also propose the Dynamic Group-wise Shift (DGW-Shift) mechanism, which is a creative combination of existing ideas from TransXNet. This mechanism allows for adaptive composition of spatial kernels based on local appearance and motion, facilitating robust implicit alignment and content-aware feature refinement in both temporal and spatial modules. The paper is well-written and easy to follow, with clear explanations of the proposed methods and experimental results. The authors provide a detailed description of the experimental setup, including the use of the HyperKvasir dataset, which is a large, publicly available collection of gastrointestinal endoscopic videos. The results show that EndoNet outperforms recent baselines in terms of PSNR and SSIM, which are common metrics for evaluating video super-resolution. The ablation studies are particularly insightful, as they demonstrate the contribution of each component of the proposed method, including the Spatial RWKV Block, the Temporal RWKV Block, and the DGW-Shift mechanism. The paper also includes a comparison with several state-of-the-art methods, such as BasicVSR, BasicVSR++, RVRT, TCNet, and IART, which helps to establish the effectiveness of the proposed approach. Overall, the paper makes a significant contribution to the field by introducing a novel architecture and a creative mechanism that address the specific challenges of EVSR.

❌ Weaknesses

Despite the paper's novel contributions and promising results, several limitations and gaps in the evaluation and analysis are evident. First, the experimental validation is limited to a single dataset, the HyperKvasir dataset, which may not be sufficient to demonstrate the robustness and generalizability of the proposed method. The HyperKvasir dataset, while large and publicly available, primarily focuses on gastrointestinal endoscopic videos. To fully validate the method's effectiveness, the authors should consider evaluating it on more diverse datasets, such as those from colonoscopy, cystoscopy, and laparoscopy, which exhibit different visual characteristics and motion patterns. The lack of such a comprehensive evaluation makes it difficult to assess the method's performance in real-world clinical scenarios, where variations in lighting, tissue textures, and the presence of artifacts like blood or bubbles are common. This limitation is particularly concerning given the importance of these factors in clinical settings, and the paper would benefit from a more thorough evaluation to address these concerns. My confidence in this issue is high, as the paper explicitly states the use of the HyperKvasir dataset and does not mention any other datasets in the experimental section. Second, the paper lacks a detailed analysis of the computational complexity and efficiency of the proposed method, especially in comparison to existing approaches. The integration of the RWKV architecture and the DGW-Shift mechanism introduces computational overhead, which is a critical consideration for real-time clinical applications where computational resources are often limited. The authors should provide a breakdown of the computational cost, including FLOPs, memory usage, and runtime, and compare these metrics against the baselines. This analysis is essential to assess the feasibility of deploying the proposed method in resource-constrained environments. The absence of such an analysis makes it difficult to evaluate the practicality of the method for real-world deployment. My confidence in this issue is high, as the paper does not mention any computational metrics in the experimental section. Third, the paper does not provide a strong justification for the choice of the RWKV architecture over other potential approaches. While the authors mention that RWKV enables efficient long-range temporal modeling, they do not provide a detailed comparison with other architectures, such as recurrent neural networks (RNNs) and transformers, in the context of EVSR. A more thorough discussion of the alternatives and the reasons for choosing RWKV would strengthen the paper. The authors should explain how the recurrent weights in RWKV contribute to long-range temporal modeling and how this differs from traditional transformer-based approaches. This would help readers understand the specific advantages of RWKV for the EVSR task. My confidence in this issue is high, as the paper only briefly mentions the benefits of RWKV without a detailed comparative analysis. Fourth, the paper does not include a discussion of the limitations of the proposed method. Addressing potential shortcomings and failure cases would provide a more balanced perspective on the method's strengths and weaknesses. For example, the paper should discuss how the method performs under extreme motion or occlusion, and how the performance is affected by different types of noise or artifacts commonly found in endoscopic videos. This would help readers understand the method's limitations and identify areas for future research. My confidence in this issue is high, as the paper lacks any dedicated section or discussion on limitations. Fifth, the paper could benefit from more background information on the RWKV architecture and the DGW-Shift module, as well as the motivation for using them. While the paper cites the original RWKV and TransXNet papers, it could provide more context on how these components are particularly well-suited for the EVSR task. A clear explanation of the mathematical formulation of the DGW-Shift module, along with a visual representation, would greatly enhance the reader's understanding. My confidence in this issue is high, as the paper's descriptions of RWKV and DGW-Shift are relatively brief and lack in-depth explanations. Sixth, the paper lacks specific details about the datasets used, including the number of videos, the duration of each video, and the types of procedures performed. This information is crucial for understanding the generalizability of the proposed method. The authors should provide a more detailed description of the dataset composition to allow readers to better assess the method's performance across different clinical contexts. My confidence in this issue is high, as the paper only provides a general description of the HyperKvasir dataset without specific details. Seventh, the paper does not provide the specific formulas for the evaluation metrics used, such as PSNR and SSIM. While these metrics are common, a more detailed explanation would be beneficial, especially for readers who may not be familiar with them. The authors should also discuss the limitations of these metrics and how they relate to the clinical relevance of the results. My confidence in this issue is high, as the paper mentions the metrics but does not provide their formulas or a detailed rationale for their use. Finally, the paper lacks detailed descriptions of the comparison methods used, including the specific architectures and training procedures. This would allow for a more thorough understanding of the advantages and disadvantages of the proposed method. The authors should provide a more comprehensive comparison with state-of-the-art methods in the field of endoscopic video super-resolution, including those that utilize deformable convolutions or other advanced alignment techniques. My confidence in this issue is high, as the paper only lists the baseline methods with citations but does not provide detailed descriptions of their architectures or training procedures.

💡 Suggestions

To address the limitations identified in the paper, I recommend several concrete and actionable improvements. First, the authors should significantly expand their experimental evaluation to include multiple datasets, such as those from colonoscopy, cystoscopy, and laparoscopy. These datasets should represent a variety of endoscopic procedures and anatomical regions to demonstrate the robustness and generalizability of the proposed method. The evaluation should not only focus on quantitative metrics like PSNR and SSIM but also include qualitative assessments through visual comparisons of the super-resolved videos. This would allow for a more comprehensive understanding of the method's performance in diverse scenarios and highlight its strengths and weaknesses in different clinical contexts. The inclusion of real-world endoscopic video datasets, if possible, would further strengthen the paper's claims about practical applicability. Second, the authors should provide a detailed analysis of the computational complexity and efficiency of their proposed method. This analysis should include a breakdown of the computational cost, such as FLOPs, memory usage, and runtime, for different components of the model. The authors should also compare these metrics against existing state-of-the-art methods to demonstrate the efficiency of their approach. Furthermore, the authors should explore potential optimizations to reduce the computational overhead, such as model pruning or quantization techniques. This is particularly important for real-time clinical applications where computational resources are often limited. The authors should also discuss the trade-offs between computational cost and performance, providing guidance on how to choose the appropriate configuration for different use cases. A clear understanding of the computational requirements is crucial for the practical deployment of the proposed method. Third, the authors should provide a more robust justification for the choice of the RWKV architecture. This should include a detailed comparison of RWKV with other potential architectures, such as recurrent neural networks (RNNs) and transformers, and explain why RWKV is particularly well-suited for the EVSR task. The discussion should cover the strengths and weaknesses of each architecture, as well as a clear explanation of how RWKV addresses the specific challenges of EVSR, such as long-range temporal dependencies and computational efficiency. This would help readers understand the rationale behind the architectural choices and the unique advantages of the proposed method. Fourth, the authors should include a dedicated section discussing the limitations of the proposed method. This should cover potential failure cases and scenarios where the method may not perform well, such as under extreme motion or occlusion, and the impact of different types of noise or artifacts commonly found in endoscopic videos. The authors should also discuss how the method performs with varying degrees of tissue deformation and lighting conditions. This would provide a more balanced perspective and help guide future research in this area. Fifth, the authors should provide more background information on the RWKV architecture and the DGW-Shift module. This should include a detailed explanation of the mathematical formulation of the DGW-Shift mechanism, along with a visual representation, to enhance the reader's understanding. The authors should also articulate the motivation for choosing these specific modules, explaining how they address the unique challenges of EVSR, such as rapid motion, non-stationary content, and varying lighting conditions. This would help readers understand the specific advantages of the proposed method. Sixth, the authors should provide more details on the datasets used, including the number of videos, the duration of each video, and the types of procedures performed. This information is crucial for understanding the generalizability of the proposed method. The authors should also discuss the specific characteristics of the datasets, such as the presence of artifacts, variations in lighting, and the types of tissue textures, to provide a more comprehensive understanding of the experimental setup. Seventh, the authors should provide the specific formulas for the evaluation metrics used, such as PSNR and SSIM, and discuss the rationale for choosing these metrics. The authors should also discuss the limitations of these metrics and how they relate to the clinical relevance of the results. This would help readers understand the evaluation process and the significance of the reported metrics. Finally, the authors should provide more details on the comparison methods used, including the specific architectures and training procedures. This would allow for a more thorough understanding of the advantages and disadvantages of the proposed method. The authors should also consider including a discussion of the trade-offs between computational complexity and performance, as this is a critical factor in practical applications. The inclusion of ablation studies on different components of the model, such as the RWKV architecture and the DGW-Shift mechanism, would also be beneficial to understand their individual contributions to the overall performance.

❓ Questions

1. Can the authors provide more insights into the training process, including the data augmentation techniques used, the learning rate schedule, and the batch size? This information is crucial for reproducibility and for understanding the factors that contribute to the performance of the method. 2. How does the proposed method perform under different challenging conditions, such as low-light scenarios or the presence of blood and other artifacts commonly found in endoscopic videos? Are there any specific adaptations or robustness measures to handle these situations? 3. How does the Dynamic Group-wise Shift mechanism compare to other adaptive kernel composition methods in the literature? What are the specific advantages and disadvantages of the proposed mechanism in the context of EVSR? 4. What are the potential limitations of the proposed method, and in what scenarios might it not perform well? How does the method handle extreme motion, occlusion, and varying degrees of tissue deformation and lighting conditions? 5. How does the proposed method compare to other state-of-the-art methods in the field of endoscopic video super-resolution in terms of performance and computational complexity? Can the authors provide a detailed comparison of the computational cost, including FLOPs, parameter counts, and inference time, and how these metrics vary across different input resolutions and hardware platforms? 6. What are the specific clinical benefits of the proposed method, and how does it improve diagnostic accuracy and surgical guidance compared to existing methods? Can the authors provide specific examples or case studies to illustrate these benefits?

📊 Scores

Soundness:2.5
Presentation:2.5
Contribution:2.0
Rating: 4.75

AI Review from ZGCA

ZGCA Review available after:
--d --h --m --s

📋 AI Review from ZGCA will be automatically processed

📋 Summary

The paper introduces EndoNet, an endoscopic video super-resolution (EVSR) framework that combines RWKV-based spatial and temporal modules with a Dynamic Group-wise Shift (DGW-Shift) mechanism for content-aware, implicit alignment. The Spatial RWKV Block enhances intra-frame structure via a Bi-WKV attention (Eq. 7) and DGW-Shift-learned depthwise kernels (Eqs. 2–4), while the Temporal RWKV Block models long-range inter-frame dependencies with linear complexity, also augmented by DGW-Shift (Section 3.3). The overall pipeline extracts features per frame (Section 3.1), applies spatial and temporal RWKV processing, and reconstructs high-resolution frames using learnable upsampling. Experiments on HyperKvasir with synthetic bicubic downsampling report small but consistent PSNR gains over CNN- and Transformer-based baselines (Table 1) and ablations demonstrate additive improvements from the spatial RWKV, temporal RWKV, and DGW-Shift components (Table 2).

✅ Strengths

  • Novel application of RWKV to EVSR: to the best of the authors’ claim, this is the first EVSR model leveraging RWKV for efficient long-range temporal modeling (Section 1; Sections 3.1–3.3).
  • Content-adaptive DGW-Shift integrated in both spatial and temporal modules to implicitly handle non-rigid motion and highlights without explicit flow/deformable alignment (Section 3.2, Eqs. 2–4; Section 3.3).
  • Clear architectural motivation for linear-time sequence modeling in long surgical videos, avoiding quadratic attention costs (Sections 1 and 3.3).
  • Systematic ablation demonstrating incremental contributions of Spatial RWKV, Temporal RWKV, and DGW-Shift (Table 2, Sections 4.4.1–4.4.2), with consistent PSNR/SSIM improvements.
  • Method section provides concrete equations for the spatial mix and channel mix layers and the Bi-WKV computation (Eqs. 5–11), aiding technical understanding.

❌ Weaknesses

  • Evaluation limited to synthetic degradations on HyperKvasir; no experiments on real clinical videos, despite the stated domain challenges (Abstract; Section 4.2; Section 5). This makes clinical relevance claims premature.
  • Incremental quantitative gains over strong baselines: +0.16 dB PSNR vs BasicVSR++ and lower SSIM than BasicVSR++ (Table 1). No statistical significance (e.g., confidence intervals) to assess robustness of small margins (Section 4.2).
  • Reproducibility and clarity issues: conflicting optimizer statements (AdamW in Section 4.1 vs Adam later in the same paragraph); conflicting use of normalization (LayerNorm in method vs 'batch normalization is applied' in Section 4.1); conflicting statements about augmentation (augmentation in Section 3.4 vs 'no additional data augmentation' in Section 4.2).
  • Backbone and citation ambiguity: feature extractor cited as ConvNeXt but referenced as Goodfellow et al. (2016) (Section 3.1), which appears incorrect; architectural details (scales, channels, tubelet shapes, values of K, G, r in DGW-Shift) are underspecified.
  • No runtime, parameter count, or FLOPs analysis, despite efficiency claims for RWKV; no wall-clock inference speed on typical resolutions/sequences (Sections 1 and 3.3 claim efficiency).
  • No temporal consistency metrics (e.g., tOF, tLPIPS) or perceptual metrics (e.g., LPIPS) and no clinical/reader study; evaluation limited to PSNR/SSIM.
  • Scope of comparison: it is not fully clear if baselines were re-trained under identical settings with matched degradation and training schedules; hyperparameters and training windows per method are not detailed (Section 4.1).

❓ Questions

  • Real-world validation: Can you provide results on real clinical endoscopic videos (even without HR ground truth) using no-reference metrics or human reader studies to assess visual quality and temporal stability?
  • Efficiency: What are parameter counts, FLOPs, and inference speed (fps) for your model compared to BasicVSR++, RVRT, and others at common resolutions and sequence lengths? How does memory scale with T?
  • DGW-Shift details: What are the specific values for K (kernel spatial size), G (number of groups), and reduction ratio r? How is DGW-Shift extended to the temporal module (temporal kernel structure, gating, and implementation details)?
  • Ablation scope: Can you include an ablation replacing RWKV with a standard transformer or a GRU/LSTM to isolate the benefit of RWKV’s recurrence and linear complexity on both accuracy and speed?
  • Optimizer and normalization: Please clarify the discrepancy in Section 4.1 (AdamW vs Adam) and normalization (LayerNorm in blocks vs 'batch normalization is applied'). Which were used in the final results?
  • Data and augmentation: Section 3.4 notes augmentation for realism, while Section 4.2 says 'no additional data augmentation'. Which is correct for the reported results? If augmentation is used, please specify types and probabilities.
  • Training seeds/hardware: What random seeds, GPUs/TPUs, and training time were used? Are results averaged across runs? Please report variance (e.g., std or CI) for PSNR/SSIM.
  • Degradation model: Are all LR inputs generated by bicubic downsampling only? Have you tried more realistic endoscopic degradations (e.g., noise, blur, illumination changes, specular corruption) and does EndoNet maintain its advantage?
  • Temporal metrics: Can you report temporal consistency metrics (e.g., tOF, tLPIPS) to support claims about stability and alignment-free design?
  • Backbone and citations: Please clarify the backbone choice and correct the citation (ConvNeXt is not Goodfellow et al., 2016). Provide an architectural diagram and layer/channel specifications to enable reproduction.

⚠️ Limitations

  • Domain gap: Training and evaluation on synthetic degradations may not capture complex real endoscopic phenomena (specularities, smoke, occlusions, non-Lambertian effects), risking reduced performance in clinical settings (Section 5).
  • Resource footprint: DGW-Shift introduces additional parameters and compute; without profiling, deployment feasibility on surgical systems remains unclear (Section 5).
  • Potential failure modes: Content-adaptive kernels could oversharpen or hallucinate fine textures in ambiguous regions, which may be risky for clinical interpretation; safeguards and uncertainty estimates are not discussed.
  • Generalization: No analysis of robustness to different scopes, patients, or imaging systems; no domain adaptation experiments despite stating this as future work (Section 5).
  • Limited metrics: Absence of temporal/perceptual/clinical metrics limits assessment of downstream utility beyond PSNR/SSIM.

🖼️ Image Evaluation

Cross‑Modal Consistency: 34/50

Textual Logical Soundness: 18/30

Visual Aesthetics & Clarity: 11/20

Overall Score: 63/100

Detailed Evaluation (≤500 words):

1. Cross‑Modal Consistency

• Major 1: Efficiency claims lack supporting runtime/FLOPs/params. Evidence: Abstract “efficient long-range…,” Sec 6 “computational efficiency” but no figure/table reporting cost.

• Major 2: SSIM claim in Conclusions contradicts Table 1. Evidence: Sec 6 “delivers higher PSNR and SSIM” vs Table 1 SSIM: 0.904 (BasicVSR++), 0.899 (Ours).

• Major 3: Visual evidence not verifiable due to illegible Fig. 1 at print size (tiny metrics/labels). Evidence: Fig. 1 (provided resolution renders PSNR/SSIM text unreadable).

• Minor 1: Fig. 1 lacks in-figure column labels (LR, M1–M4, Ours, GT), forcing reliance on caption. Evidence: Fig. 1 (no embedded legend/labels).

• Minor 2: Method says “synthetic endoscopic video dataset” (Abstract), but Experiments are only HyperKvasir. Evidence: Abstract vs Sec 4 (HyperKvasir only).

2. Text Logic

• Major 1: Optimizer inconsistency (AdamW vs Adam). Evidence: Sec 4.1 “AdamW…,” then “The model is optimized using Adam…”.

• Major 2: Data augmentation inconsistency. Evidence: Sec 3.4 “data augmentation…,” vs Sec 4.2 “no additional data augmentation is applied”.

• Major 3: “Extensive experiments” but single dataset and no statistical tests. Evidence: Abstract “Extensive experiments…,” Sec 4 uses only HyperKvasir.

• Minor 1: Misattributed backbone reference. Evidence: Sec 3.1 “ConvNeXt Goodfellow et al. (2016)” (Goodfellow is a textbook, not ConvNeXt).

• Minor 2: Section title duplication. Evidence: Sec 4.4.1 “QUANTITATIVE COMPARISON” repeats Sec 4.3 title.

3. Figure Quality

• Major 1: Fig. 1 illegible at ≈100% print size; critical metrics/annotations unreadable; blocks verification of qualitative claims. Evidence: Fig. 1 (tiny overlaid PSNR/SSIM and ROI text).

• Minor 1: Figure‑alone test fails—no embedded legend mapping columns to methods; needs call‑outs/labels. Evidence: Fig. 1 (no column headers inside the image).

Key strengths:

  • Clear motivation for EVSR and well‑articulated challenges in endoscopy.
  • Method combines RWKV with a Dynamic Group‑wise Shift; ablation table shows consistent gains.
  • Equations specify spatial/temporal mixing and DGW‑Shift weighting pipeline.

Key weaknesses:

  • Central efficiency claims lack any computational evidence.
  • Inconsistencies (optimizer, augmentation) reduce reproducibility.
  • Fig. 1 is not legible or self‑contained; qualitative support is weakened.
  • Overclaim in Conclusions (SSIM) contradicts Table 1; “extensive” evaluation limited to one dataset.

Recommended fixes (high impact):

  • Add a table with params/FLOPs, inference time vs sequence length, and memory.
  • Resolve optimizer and augmentation contradictions; detail exact training protocol.
  • Replace Fig. 1 with high‑resolution, add in‑figure column labels and readable metrics.
  • Soften Conclusions (SSIM) or provide evidence; consider adding another dataset or clinical videos.

📊 Scores

Originality:3
Quality:2
Clarity:2
Significance:2
Soundness:2
Presentation:2
Contribution:2
Rating: 4

AI Review from SafeReviewer

SafeReviewer Review available after:
--d --h --m --s

📋 AI Review from SafeReviewer will be automatically processed

📋 Summary

The paper introduces EndoNet, a novel framework for endoscopic video super-resolution (EVSR) that combines the Receptance Weighted Key Value (RWKV) architecture and a Dynamic Group-wise Shift (DGW-Shift) mechanism. The RWKV architecture is used for efficient long-range temporal modeling, while the DGW-Shift mechanism adaptively composes spatial kernels based on local appearance and motion, enabling robust implicit alignment and detail restoration without explicit motion estimation. The authors claim that EndoNet achieves a strong balance between global context modeling and local adaptability, maintaining small yet stable advantages over recent CNN- and transformer-based baselines in quantitative comparisons. The paper's core contributions lie in the innovative application of RWKV to the EVSR task and the introduction of the DGW-Shift mechanism. The method is evaluated on the HyperKvasir dataset, and the results show that EndoNet outperforms several baselines in terms of PSNR and SSIM. However, the paper's significance is somewhat diminished by the limited scope of the experimental evaluation and the lack of detailed analysis of computational efficiency and real-world applicability. Despite these limitations, the paper provides a promising direction for future research in EVSR by exploring the potential of RWKV and adaptive kernel composition.

✅ Strengths

One of the key strengths of this paper is the innovative application of the RWKV architecture to the EVSR task. The RWKV architecture, originally developed for natural language processing, is adapted to handle the long-range temporal dependencies inherent in endoscopic videos. This adaptation is particularly noteworthy as it addresses the computational inefficiencies of traditional transformer-based models, which often struggle with the quadratic complexity in long sequences. The introduction of the Dynamic Group-wise Shift (DGW-Shift) mechanism is another significant contribution. This mechanism allows the model to adaptively compose spatial kernels based on local appearance and motion, enabling robust implicit alignment and detail restoration without explicit motion estimation. The paper's method section provides a clear and detailed description of the proposed framework, including the mathematical formulations and the integration of the RWKV and DGW-Shift components. The authors also conduct extensive experiments on the HyperKvasir dataset, which is a relevant and challenging dataset for EVSR. The quantitative results show that EndoNet outperforms several recent baselines in terms of PSNR and SSIM, demonstrating the effectiveness of the proposed approach. The ablation studies further validate the contributions of the spatial and temporal RWKV blocks and the DGW-Shift mechanism, providing insights into the model's design and performance. Overall, the paper's technical innovations and empirical achievements make it a valuable contribution to the field of EVSR.

❌ Weaknesses

Despite the paper's strengths, several weaknesses and limitations need to be addressed to enhance its overall quality and impact. One of the most significant concerns is the clarity of the paper, particularly in the method section. The description of the RWKV architecture and its integration into the EVSR framework is dense and may be challenging for readers unfamiliar with RWKV. For instance, the paper introduces numerous mathematical notations and concepts without providing sufficient context or intuitive explanations. The equations, while mathematically sound, lack accompanying diagrams or visual aids that could help readers better understand the flow of information and the role of each component. This issue is particularly evident in Section 3.2, where the DGW-Shift mechanism is described. The lack of visual aids and the rapid introduction of technical details make it difficult to follow the proposed method, potentially limiting the paper's accessibility and impact.

Another critical weakness is the limited experimental evaluation. The paper primarily evaluates EndoNet on the HyperKvasir dataset, which, while relevant, does not provide a comprehensive assessment of the model's generalization capabilities. The inclusion of additional datasets, such as Endo-Vid and Kvasir-V2, would strengthen the evaluation and demonstrate the robustness of the proposed method across different endoscopic scenarios. Furthermore, the paper does not provide a detailed analysis of the model's performance on real-world endoscopic videos. The reliance on synthetic data for evaluation raises concerns about the practical applicability of the method, as real-world videos often contain complex artifacts and variations that are not present in synthetic data. The authors should conduct experiments on real clinical data to validate the model's performance in realistic scenarios.

The paper also lacks a thorough analysis of computational efficiency, which is a crucial aspect for real-time applications like EVSR. While the authors mention the theoretical computational advantages of RWKV, they do not provide concrete metrics such as FLOPs, model parameters, or inference time. This omission makes it difficult to assess the practical feasibility of the proposed method, especially in resource-constrained environments. The inclusion of these metrics would provide a more comprehensive comparison with existing methods and help readers understand the trade-offs between performance and computational cost.

The ablation studies, while present, could be more detailed and informative. The current ablation study in Table 2 shows the impact of removing the Spatial RWKV Block, Temporal RWKV Block, and DGW-Shift mechanism on PSNR and SSIM. However, the paper does not provide a deeper analysis of the specific contributions of each component, such as the impact of different configurations of the RWKV state size or the number of DGW-Shift groups. A more granular ablation study would help readers better understand the model's design and the importance of each component.

Additionally, the paper's writing style could be improved. The method section is overly dense and technical, which may hinder readability. The authors should consider simplifying the language and providing more intuitive explanations of the proposed method. The use of visual aids, such as diagrams and flowcharts, would also enhance the clarity of the paper. The current presentation of equations and technical details is not sufficiently supported by visual representations, making it challenging for readers to grasp the core concepts.

Finally, the paper lacks a dedicated limitations section. While the authors briefly mention some limitations in the discussion section, a more thorough and explicit discussion of the model's limitations and potential failure cases would provide a more balanced and realistic assessment of the proposed method. This section should address the challenges of handling extreme non-rigid deformations, severe occlusions, and the computational cost of the model. The inclusion of such a section would demonstrate the authors' awareness of the model's shortcomings and guide future research in this area.

💡 Suggestions

To address the identified weaknesses, I recommend several concrete and actionable improvements. First, the authors should enhance the clarity of the paper by providing more intuitive explanations of the RWKV architecture and the DGW-Shift mechanism. This could involve including diagrams or flowcharts that visually represent the flow of information and the role of each component. The equations should be accompanied by clear descriptions of the variables and their significance, making the method more accessible to a broader audience. Additionally, the authors should consider simplifying the language in the method section to improve readability.

Second, the experimental evaluation should be expanded to include additional datasets, such as Endo-Vid and Kvasir-V2. This would provide a more comprehensive assessment of the model's generalization capabilities and demonstrate its robustness across different endoscopic scenarios. The authors should also conduct experiments on real-world endoscopic videos to validate the model's performance in realistic clinical settings. This could involve a smaller-scale study with a focus on visual quality and clinical relevance, as well as a discussion of the challenges and potential solutions for handling real-world artifacts and variations.

Third, the paper should include a detailed analysis of computational efficiency. The authors should provide metrics such as FLOPs, model parameters, and inference time for EndoNet and compare them with existing methods. This analysis should be conducted on a standard hardware setup to ensure reproducibility and provide a clear understanding of the practical feasibility of the proposed method. The authors should also discuss the trade-offs between performance and computational cost, which is crucial for real-time applications like EVSR.

Fourth, the ablation studies should be more detailed and informative. The authors should explore the impact of different configurations of the RWKV state size, the number of DGW-Shift groups, and other architectural parameters. This would help readers understand the specific contributions of each component and the optimal settings for the model. The ablation study should also include a comparison of the proposed method with and without the DGW-Shift mechanism to demonstrate its effectiveness in handling complex motion and occlusions.

Finally, the paper should include a dedicated limitations section that explicitly discusses the model's shortcomings and potential failure cases. This section should address the challenges of handling extreme non-rigid deformations, severe occlusions, and the computational cost of the model. The authors should also discuss the potential for further improvements and future research directions, such as exploring domain adaptation techniques to bridge the gap between synthetic and real-world data. By addressing these weaknesses, the paper can provide a more comprehensive and balanced evaluation of the proposed method, enhancing its overall quality and impact.

❓ Questions

1. Could the authors provide a more detailed explanation of the RWKV architecture and its specific adaptations for the EVSR task? For instance, how are the state representations initialized, and what are the specific configurations of the RWKV layers used in the model?

2. How does the Dynamic Group-wise Shift (DGW-Shift) mechanism specifically address the challenges of non-rigid tissue deformation and rapid camera motion in endoscopic videos? Could the authors provide visual examples or case studies that demonstrate the effectiveness of DGW-Shift in these scenarios?

3. What is the computational cost of the proposed method in terms of FLOPs, model parameters, and inference time? How does this compare to existing CNN- and transformer-based methods for EVSR, and what are the implications for real-time processing?

4. Could the authors conduct experiments on additional datasets, such as Endo-Vid and Kvasir-V2, to evaluate the generalization capabilities of the proposed method? How does the model perform on these datasets, and what are the key differences compared to the HyperKvasir dataset?

5. How does the proposed method handle extreme cases of non-rigid deformation and severe occlusions by surgical tools? Could the authors provide examples or analysis of the model's performance in these challenging scenarios, and discuss any potential limitations or failure cases?

📊 Scores

Soundness:2.25
Presentation:2.25
Contribution:2.25
Rating: 4.75

Keywords

Click the button to extract keywords

Insights

Click the button to extract insights
Version Management
Version 1
Citation Tools

📝 Cite This Paper