EndoDGS: Degradation-Decoupled Gaussian Splatting for Endoscopic Novel-View Reconstruction

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript presents an endoscopic 3D Gaussian Splatting framework for degradation-decoupled reconstruction under challenging endoscopic imaging conditions. The proposed framework is technically well organized, and the experimental results on both synthetic and real datasets demonstrate the effectiveness of the proposed approach. The topic is relevant to endoscopic 3D reconstruction and neural rendering. However, several aspects of the presentation and experimental analysis could be further improved to enhance the clarity and overall quality of the manuscript. The following comments are provided for the authors' consideration.

1.

The abstract is well organized, with a clear motivation and a coherent presentation of the proposed framework. However, the method description is relatively dense, and the main technical contribution is not sufficiently emphasized. A more concise presentation with clearer emphasis on the core innovation would further improve the readability of the abstract.

2.

Figure 1 effectively illustrates the motivation of the proposed framework. However, the black text blends with the dark background of the corresponding text boxes, making labels such as "Trajectory-guided Replay" less visually distinguishable. Improving the visual contrast would enhance the readability of the figure.

3.

The introduction provides sufficient background and motivation for the proposed work. However, it is relatively lengthy and contains some repetitive descriptions. In addition, several method-specific details may be more appropriately presented in the methodology section. The novelty with respect to existing endoscopic Gaussian-based reconstruction methods could also be emphasized more clearly.

4.

The related work section provides a comprehensive review of existing studies. However, the comparison with recent endoscopic Gaussian-based reconstruction methods remains relatively general. A more specific discussion of the differences between the proposed framework and existing approaches would better highlight the novelty of the proposed method.

5.

Figure 2 provides a comprehensive overview of the proposed framework. However, the workflow contains many visual components and annotations, making the main processing pipeline less immediately identifiable. A clearer organization of the visual elements would improve the readability of the figure.

6.

The system overview clearly presents the overall reconstruction pipeline. However, Equations (2)–(6) mainly describe the sequential processing flow of the proposed framework. A brief summary highlighting the role of each stage before introducing the mathematical formulations would improve the readability of this section.

7.

The methodology is comprehensive and systematically introduces the proposed modules. However, several subsections, particularly LAEM and SDC, contain lengthy descriptions before the mathematical formulations. A more concise presentation would improve readability while preserving the technical completeness of the methodology.

8.

The experimental settings are described in sufficient detail. However, the reasons for selecting several important hyperparameters, such as the loss weights and appearance modulation settings, are not discussed. Providing a brief explanation for these design choices would improve the transparency of the experimental setup.

9.

The benchmark experiments demonstrate consistent improvements over the compared methods. However, the quantitative analysis mainly focuses on describing the observed performance differences. A more in-depth discussion explaining why the proposed degradation-decoupled design leads to the reported improvements under different imaging conditions would further strengthen the experimental analysis.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The current manuscript describes the methods of degradation-decoupled Gaussian splatting in application to endoscopic novel-view reconstruction analysis. Clearly, the material is important and the new developments in the image-recognition are important per se, but the current manuscript does require substantial modifications prior to acceptance.

 

First, i found the language being not particularly well polished: the authors should focus on this aspect in the revised version.

 

In Secs. 3.3 and 3.4 where the technical details of the image-analysis algorithms are presented, the authors should possible dissolve this very technical material with some illustrative examples showing which values these and those quantities and quantifiers would take for some sort of simple/typical artificial or synthetically generated images. This way, when presenting then the results for real images, the reader will instantly understand how the values for the same parameters for REAL images are different from those obtained for idealized or artificial ones.

 

The authors do present something similar for a set of synthetic images in figs. 5 and 6, but the discussion of the obtained results should be improved and extended considerably.

 

​Presenting after such a long main part and after the depiction of synthetic images only a simple fig. 7 with real images seems somewhat insufficient. Here, the experienced readers would like to see a set of dedicated images for each of the most important types of abnormalities detected during typical colonoscopy studies and a detailed analysis of each sort of such pathologies with the help of the current method, in comparison to other methods of the image-analysis and pattern-recognition available on the modern computer market.

 

The conclusions should certainly be extended.

 

Lastly, the bibliography is still not detailed and broad enough, in my opinion. Here, the authors can e.g. discuss some recent models of image-pattern- and feature-detection-analysis from close and other disciplines, and in particularly those working on the dynamical data sets, instead of on the static images as most of such algorithms, including the current one. One of the references which can be mentioned in this context is the recent and relevant study [DOI: https://doi.org/10.1103/PhysRevResearch.5.043129].

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Considerable improvements in all sections of the manuscript: the material can be accepted.