Text Promptable Surgical Instrument Segmentation with Vision-Language Models

In this paper, we propose a novel text promptable surgical instrument segmentation approach to overcome challenges associated with diversity and differentiation of surgical instruments in minimally invasive surgeries. We redefine the task as text promptable, thereby enabling a more nuanced comprehension of surgical instruments and adaptability to new instrument types. Inspired by recent advancements in vision-language models, we leverage pretrained image and text encoders as our model backbone and design a text promptable mask decoder consisting of attention- and convolution-based prompting schemes for surgical instrument segmentation prediction. Our model leverages multiple text prompts for each surgical instrument through a new mixture of prompts mechanism, resulting in enhanced segmentation performance. Additionally, we introduce a hard instrument area reinforcement module to improve image feature comprehension and segmentation precision. Extensive experiments on several surgical instrument segmentation datasets demonstrate our model's superior performance and promising generalization capability. To our knowledge, this is the first implementation of a promptable approach to surgical instrument segmentation, offering significant potential for practical application in the field of robotic-assisted surgery.

Paper

Similar papers

Peer review

Reviewer fFdZ5/10 · confidence 4/52023-07-01

Summary

This manuscript presents a clip-assisted semantic image segmentation method for surgical instruments. In terms of methodology, the proposed work can be viewed as an adaption of CLIPSeg to a domain-specific problem. Compared with CLIPSeg, a mixture of prompt strategy is used for augmenting text prompt information. Hard sample mining is also employed to further improve segmentation performance. The image segmentation network architecture has also been optimized. The proposed method is evaluated on two public surgical instrument segmentation datasets and is shown to outperform some fully-supervised and clip-assisted methods. **Commented after rebuttal**: I would like to thank the authors, other reviewers, and the ACs. I have carefully read through the rebuttals, and the comments from other reviewers. The discussion with the authors is constructive. Most of my raised concerns have been properly addressed.

Strengths

The proposed method yields improved performance on two surgical instrument segmentation datasets, compared with some existing fully-supervised/clip-assisted methods. Employing text prompt brings additional flexibility in terms of the categories of segmentation targets. This is a major advantage of clip-assisted segmentation over conventional fully-supervised segmentation. The proposed method is introduced in sufficient detail and can be easy to follow.

Weaknesses

The take-home information for readers may be limited/unclear. Despite that domain-specific adaptions on top of vanilla CLIPSeg leads to improved performance on surgical instrument segmentation, these contributions themselves may not be of sufficient interest to readers: hard sample mining, feature pyramid, generating multiple text prompts are already the common practice and should have been well-known for the community. The authors are encouraged to highlight their key take-home information to readers, and argue how the proposed work brings new knowledge. In the abstract and introduction, the authors highlight improved flexibility when dealing with new segmentation targets of clip-assisted segmentation models. However, this point is not sufficiently evaluated and discussed in the experiments: how does CLIPSeg and CRIS that are also clip-driven work in this context? The authors may also want to discuss/comment on a closely-related existing work [1]. 1. CLIP-Driven Universal Model for Organ Segmentation and Tumor Detection

Questions

Judging by the ablation study (table 3 - 6), the contributions of individual components, except for the text encoder, seem to be quite marginal (the ablated models seem to be already quite strong). Therefore, from the authors' point of view, which factor accounts for most of the improved performance over CLIPSeg and CRIS? Or if there are other components that make a difference? Does the proposed method employ a stronger segmentation backbone or better training technique? The authors are encouraged to provide more details about the compared methods to avoid confusion. From the authors' perspective, what is the core take-home information to the readers?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Automated instrument image segmentation may affect the outcome of a surgery, which may lead to real-world impact to the clinicians and the patient. The authors are encouraged to discuss the potential impact of their work for patients, clinicians, device manufacturers, and the society. The authors are also encouraged to discuss if the proposed method exacerbate/mitigate potential risks and biases in image segmentation.

Reviewer 9oei5/10 · confidence 4/52023-07-03

Summary

This paper proposes a novel text promptable surgical instrument segmentation approach to overcome challenges associated with the diversity and differentiation of surgical instruments by using the large CLIP model and a text promptable mask decoder. The experiments show the effectiveness of the proposed method on surgical instrument segmentation. However, I am very concerned with the novelty of the proposed modules due to the limited situations.

Strengths

1, The paper is well-written and easy to understand. 2, The experiments show the effectiveness of the proposed modules.

Weaknesses

1, There are too few novelties due to most of the proposed modules being explored in the traditional segmentation tasks. 2, The dataset is too limited to prove the effectiveness of the proposed modules. Furthermore, there is no specific module designed for surgical instrument segmentation. 3, Lacking the comparison of inference time.

Questions

1, How about comparing the proposed method and the Segment Anything model? 2, What’s the difficulty of surgical instrument segmentation? I understand that there might be a lack of some segmentation datasets leading to overfitting.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

N/A

Reviewer Rt9K7/10 · confidence 4/52023-07-06

Summary

This paper presents a text prompt-based surgical instrument segmentation method, which is more scalable to handle the diverse targets in endoscopy videos.

Strengths

- originality: this is the first work to use text prompt for surgical instrument segmentation - quality: the performance of the proposed methods surpassed previous methods - clarity: the paper is well organized - significance: the method can handle unseen targets, which is desired in clinical scenarios.

Weaknesses

- EndoVis challenge is organized every year. However, this paper validated the method on old datasets (18-19). Why not use the latest dataset and compare it to the challenge winning solutions? e.g., 21-22 http://opencas.dkfz.de/endovis/challenges/2022/ - Reference formats are not consistent.

Questions

- Does the hard instrument area identified by the model align with the surgeons' perspective? - Since the best performance in Table 1-2 still has a large room for further improvement. What are the typical failure modes? What's the potential reason for the model to generate such segmentation errors?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

A more comment task is to segment both instruments and tissues. It would be great to validate the method in this setting. Here is a public dataset https://www.kaggle.com/datasets/newslab/cholecseg8k

Reviewer NUuZ7/10 · confidence 4/52023-07-08

Summary

The paper introduces a novel approach for surgical instrument segmentation in minimally invasive surgeries. By leveraging text prompts and vision-language models, the proposed method achieves improved segmentation performance. The approach shows promise for practical use in robotic-assisted surgery.

Strengths

The present work contributes with an innovative and effective approach for text promptable surgical instrument segmentation in minimally invasive surgeries. This paper presents a meticulous study of previous work, which is important in the development of the present work. Also, the technical aspects are clearly explained and have also been evaluated using the correct metrics. Another strength of this paper is the introduction of a mixture of prompts mechanism. By leveraging multiple text prompts for each surgical instrument, the authors enhance the segmentation performance of their model. The experimental evaluation of the proposed model on EndoVis2017 and EndoVis2018 datasets demonstrates its superior performance compared with other works and promising generalization capability. In summary, the work is an interesting application of deep learning in the medical area, and it also has a remarkable novelty.

Weaknesses

Regarding the ablation study, it would be good if the authors could explain why they chose 448x448 as the image size. Aren't some details lost using this size? It would be good if the authors could give more details about the dataset, i.e. the average duration of each video for example. It would be very positive to also include more datasets, such as EndoVis2019, EndoVis2020 or EndoVis2021.

Questions

Do you plan to include more datasets such as EndoVis2019, EndoVis2020 or EndoVis2021?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

limitations are not mentioned, authors should include the limitations in the paper.

Area Chair PXWr2023-08-19

This is a friendly reminder from the AC that you need to respond to the rebuttal, since the authors spent quite a lot of time preparing the rebuttal.

Reviewer wnjR6/10 · confidence 4/52023-07-27

Summary

This paper introduces a novel idea of utilizing text prompts and vision-language models to make surgical instrument segmentation more flexible and robust to diversity. The proposed method and custom modules achieve strong results on two endoscopic datasets.

Strengths

1. The paper tackles an important problem in surgical instrument segmentation, which aims to enhance robot-assisted surgery systems. The key idea of using text prompts to improve generalization and adaptability to new instruments is novel. 2. The method is technically sound, leveraging recent advances in vision-language models like CLIP. The image and text encoder setup makes sense. The text promptable mask decoder uses attention and convolution schemes nicely for decoding. 3. Several custom modules are proposed to boost segmentation performance: 1) Mixture of prompts leverages multiple prompts effectively. 2) Hard instrument area reinforcement focuses on challenging regions. 4. Comprehensive experiments on two datasets demonstrate superior performance over state-of-the-art methods. The cross-dataset generalization results are promising. The ablation studies validate the efficacy of individual model components like multi-scale feature extraction, mixture of prompts, and hard area reinforcement.

Weaknesses

1. The problem definition and goal can be further sharpened. How does text-based prompting specifically help with increasing instrument variety and subtle inter-class differences? This needs more elaboration upfront. 2. Some architectural details are unclear - like how exactly text features are integrated into the convolutional prompting scheme. More implementation specifics will help reproducibility. 3. The computational complexity and inference speed are not analyzed. This could be important for practical usage. 4. More in-depth experimentation on real-world surgical videos and systems would be preferred to further demonstrate applicability.

Questions

Here are some suggestions for authors to consider: 1. Explain how text prompting helps with adaptation to new instruments and resolving inter-class confusions, with examples. 2. The method sections explain each component logically but can be more crisp and coherent in places. Authors could provide more implementation specifics for text feature integration into convolutional prompting and gating network. 3. Consider reporting computational complexity and inference speeds. How does it compare with prior arts? 4. Consider detecting and segmenting novel instruments not seen during training by using only text prompts. Typos: 1. line 58: "launguage" should be "language". 2. Mixed use of "visual-textual" and "visual-textural" at multiple places. Authors should also consider discussing the limitations, such as those listed below, in the paper.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

1. The method has only been evaluated on two datasets with limited surgical scenarios. Performance on more diverse real-world data is unclear. 2. It relies on high-quality textual prompts, which may not always be available or easy to construct in practice. 3. The requirement of retraining with new text prompts for new instruments reduces adaptability.

Reviewer Rt9K2023-08-14

Thanks for your detailed explanation very much. My major concern has been addressed. One minor question, if the paper is accepted, could you please promise to make the complete training and inference code publicly available by Dec.?

Authorsrebuttal2023-08-14

Thank you very much for your reply. If our paper is accepted, we promise to release the code by December.

Reviewer 9oei2023-08-14

Hi authors, Thanks for your hard work on the rebuttal. Overall, I am familiar with general segmentation tasks instead of surgical segmentation. It is challenging for me to make the proper justification in this area. Therefore, I have carefully read other reviewers' comments and see their appreciation of this work. Finally, I change my original score to borderline acceptance. I cannot give higher scores due to that I still lack specific knowledge. Hope the SPCs and ACs can see more comments from other reviewers.

Authorsrebuttal2023-08-14

Many thanks for acknowledging our rebuttal and paper, we will continuely improve our paper according to your comments in the revised version.

Reviewer fFdZ2023-08-15

Hi authors, Thank you so much for the clarification. Most of my concerns are properly addressed and I am very pleased to see the proposed method yielding improved performance compared with existing prompt-driven methods with large margins on both IID and OOD settings. I am also pleased to see the in-depth analysis about the potential impacts and biases, which comprises responsible research and innovation. Just two minor comments: 1. I am still a bit missing about the core take-home information. Despite the authors' claim to be the first to apply prompt-driven segmentation to surgical instrument segmentation, unless the drastic differences between segmenting surgical instruments versus segmenting radiological images/RGB images are adequately explained, I would not be 100% convinced that itself comprises a major contribution in the context of NeurIPS. 2. Still, the improved segmentation performance attribute to a synergy among multiple components, instead of the prompt-driven mechanism alone. The authors are expected to make this point very clear to readers, and emphasize/argue a) why a synergy among these components is crucial for the targeted task; b) why the proposed synergy is applicable to other domains that are relevant to the readers of NeurIPS (e.g., segmenting RGB and radiological images).

Authorsrebuttal2023-08-16

Thank you very much for acknowledging our rebuttal and paper. Below, we provide answers to your newly raised comments: ## Q1 Thanks for your question. Below we summarize the challenges lying in surgical images: - In surgical images, it is common that instruments may exhibit significant similarity, necessitating the model that can distinguish their subtle differences for accurate segmentation. - As surgeries progress, surgical instruments might cut, suture, or otherwise manipulate tissues, altering their shapes. This can lead to tissues obscuring the instruments and potentially making some parts of the tissues even resemble the instruments. - In surgical settings, the use of diminutive endoscopic devices with small lenses inherently constrains the imaging quality due to hardware limitations. - The areas into which the endoscope is inserted, like the gastrointestinal tract, are continuously moving, which can add complexities to the segmentation of surgical instruments. Given above, on one hand, when comparing with those objects in natural scenes in RGB images which often have distinct and rigid boundaries, the continuous morphological changes of instruments and tissues during surgeries make instrument segmentation easily affected by tissue occlusions and variations in illumination, etc. On the other hand, when dealing with radiological images, different segmentation challenges emerge: the presence of various imaging modalities, intrinsic noise, limited contrast, and the potential for artefacts collectively introduce complexity to the precise segmentation in radiological images. Overall, we firmly believe that research on surgical instrument segmentation is both challenging and meaningful. Our methods have achieved significant improvements over text-promptable methods (e.g., CRIS, CLIPSeg) developed on natural images. We follow a problem-driven approach, in response to the aforementioned challenges in surgical images, we introduce the mixture of prompts module to address instrument similarity by integrating detailed text descriptions; moreover, the hard instrument area reinforcement module further amplifies the model's precise segmentation performance, especially in tricky regions where distinctions between instruments and tissues become blurred. Finally, although our method is designed for surgical instrument segmentation, it should also have the merits in natural image segmentation, especially when encountering challenges similar to those in surgical contexts. For instance, when undertaking fine-grained segmentation or in the presence of heavy occlusions. Therefore, our method holds potential for broader applications in computer vision tasks, and NeurIPS would be an excellent venue to showcase our work. ## Q2 Thanks for this question. First, we emphasize that our method is not merely a synergy of a few components. For the text-promptable pipeline itself, there are inherent novelties instead of a simple adaptation from existing approaches. For instance, we devise a multi-scale fusion scheme for image encoder and the mask decoder integrated both attention-based and convolution-based prompting schemes to facilitate text features in guiding visual features for segmentation prediction (see Tab. 3 & 4 in our paper for their improvements). Building upon the text-promptable method, we introduce the mixture of prompts, substantially enhancing model performance, especially in the prediction of novel instruments (see Tab. 5 in our paper). Additionally, to overcome classification inaccuracies during segmentation, we incorporate the hard instrument area reinforcement module (see Tab. 6 in our paper). All these modules, as mentioned in the above answer, follow a problem-driven paradigm in the surgical instrument segmentation domain. Next, although our modules have been primarily validated in surgical instrument segmentation, we believe they can offer insights or potential benefits for segmentation tasks in other domains (e.g., RGB and radiological images), especially when facing challenges akin to those found in our surgical images. For example, in tasks such as fine-grained segmentation of natural images, our mixture of prompts module could be effectively employed and for scenarios with heavy occlusions, our hard instrument area reinforcement module could be particularly suited. Our approach exhibits strong adaptability, as demonstrated on the CholecSeg8k dataset where it not only segments instruments but also various tissues. Experimentally, our method surpasses the current SOTA in performance (see results in the global response). In summary, our method was borne out of a problem-driven necessity. To address these challenges, we introduce various modules which, when synergized, significantly enhance performance. We believe our approach harbors immense potential for broader applications within the computer vision domain, especially for methods centered on text prompts.

Reviewer fFdZ2023-08-17

Hi Authors, Thank you for your clarifications. I would encourage you to concisely emphasize your arguments above in the paper. These would help the readers to grasp the nature of your problem as well as the key contributions. (Still I am not fully convinced by response 2: I would argue performance improvement would also be seen when hard negative mining and feature pyramid were applied to ClipSeg.)

Authorsrebuttal2023-08-18

Thanks a lot for your response, we will surely add the relevant content in rebuttal into the final version of our paper. Regarding the Hard Instrument Area Reinforcement (HIAR) module, we agree that improvement might be observed if it is applied to CLIPSeg. As being said, our method was designed in a problem-driven manner, integrating HIAR into CLIPSeg is likely to also improve the performance in surgical instrument segmentation domain. Besides, for similar scenarios in natural images, e.g. objects are occluded, integrating HIAR into CLIPSeg might also help. Overall, these potential improvements indeed show the generalizability of our proposed module as well as its merits beyond the surgical instrument segmentation task.

Reviewer wnjR2023-08-17

Thank you for the detailed rebuttal. I appreciate the additional experimental results and analysis. Please try to incorporate the details mentioned in the rebuttal into the revised version. Since more datasets are involved, I think it will be beneficial to also include more cross-dataset results to better demonstrate the robustness of the proposed method for segmenting unseen instruments. I presume it is relatively easy to complete as no retraining is needed. In general, the authors' rebuttal resolved most of my concerns. I'm raising my rating to weak accept.

Authorsrebuttal2023-08-17

Thank you very much for recognizing the value of our rebuttal. We will include the additional experimental results from the rebuttal phase in our revised paper. We appreciate your suggestion, and we also plan to incorporate more cross-dataset validation results in the revised paper. We conducted a quick cross-dataset experiment using a model trained on EndoVis2017 and tested it on the EndoVis2019 dataset. Our method achieved DSC=0.90 and NSD=0.90, which surpasses the competition's previous best result of DSC=0.89 and NSD=0.89. This further underscores the exceptional performance of our approach. We will provide more cross-dataset experimental results in our subsequent revised paper.

Reviewer NUuZ2023-08-21

Final remarks

Thank you for your comprehensive response to my review of your paper. I appreciate the clarifications and additional details you've provided in response to the concerns I raised. Certainly, your explanations have explained better different aspects of your work. I appreciate your explanation regarding the choice of the input image size (448x448) and your consideration of the requirements imposed by the ViT-based image encoder. The provided information clears up any confusion I had regarding this matter. Thank you for providing detailed descriptions of the datasets used in your experiments. The additional information you've shared regarding EndoVis2017, EndoVis2018, EndoVis2019, and CholecSeg8K is invaluable in understanding the scope and diversity of the data you've employed. I am pleased to see that you've taken our suggestion into consideration and included experimental results on EndoVis2019 and CholecSeg8K. This expansion of your evaluation enhances the robustness and generalizability of your findings, and we believe it will strengthen the paper's contribution. Based on your detailed responses and the additional information you've provided, I am confident that your paper offers a positive contribution to the domain of text-promptable surgical instrument segmentation.

Authorsrebuttal2023-08-21

Thank you very much for your recognition. We will continue to improve our paper and incorporate the information from the rebuttal.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC