Aligning Synthetic Medical Images with Clinical Knowledge using Human Feedback

Summary

This paper proposed to use the diffusion model with expert feedback to improve the quality of synthetic medical images. The key motivation is the embedding of RLHF for medical image generation. To achieve this, there are three steps: (1) pre-trained a generation model and collected expert feedback based on several manual rules; (2) trained a model to predict the feedback scores; (3) incorporated the scores to fine-tune generative models. Extensive experiments demonstrate the effectiveness of the proposed model. Overall, this is an interesting and sound paper.

Strengths

1. RLHF-based generation model for medical knowledge incorporation. Interesting and useful. 2. Reasonable framework. The proposed model is based on the conditional diffusion models and the whole process is sound. 3. Extensive evaluation. Besides the commonly-used quality-related metrics and downstream performance, the new knowledge discovery and visual results are also impressive.

Weaknesses

1. Reproductivity. It's better to make the codes public and improve the reproductivity. 2. Manual rules. The expert feedback is highly related to the manual design rules. It's better to see the performance with different amounts of rules or with more rules. Also, please elaborate more on how to define these rules. In other words, it is not clear why the designed rules can indicate the complex image quality and small image differences. 3. Downstream performance. There is only one downstream task to denote effectiveness. More tasks should be considered. Meanwhile, why not show the performance of downstream models trained by both synthetic and real images?

Questions

See the above weaknesses.

Rating

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

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

Besides the comments on the paper's weaknesses, please also consider: 1. Even though it is a kind of application paper, please add some discussions to state the technical contributions. 2. The authors claimed such a synthetic way could be useful for rare diseases. Please provide the evidence.

Summary

This paper develops a framework that generates synthetic medical images aligned with the clinical knowledge of doctors through training a reward model based on pathologists' image plausibility annotations. The intuition for their very simple but powerful approach that depends on incorporating clinical expert knowledge into the diffusion models by using the reward model to inform a finetuning objective arises from their observation: "designing domain-specific objective functions for finetuning foundational models that ensure a generative model adheres to clinical knowledge is challenging". Their presentation is very clear, and empirical evaluations & ablations are convincing. I recommend this paper for acceptance beyond my concerns around how well this paper fits into NeurIPS venue.

Strengths

- Very clear presentation - Simple but powerful idea that brings in an idea from a different domain (LLMs) to medical imaging - Convincing empirical results and ablations

Weaknesses

My main worry about this paper is that NeurIPS might not be the best venue, given the algorithmic novelty is limited here. However, I do think this is an interesting application of integrating domain-specific knowledge to models through RLHF-style training.

Questions

- How do you think your findings might change as the diffusion models get stronger in creating more plausible images (given the rate of implausible images was higher than 80% for some cell types)? - How do you think you might generalize your method to other modalities such as interpreting MRI images where clinical plausibility criteria might be more limited? (might be harder to define the criteria beyond aliasing artifacts in certain anatomies?)

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

4 excellent

Contribution

2 fair

Limitations

Authors have not explicitly described limitations of their work.

Summary

The authors proposed to include pathologist in the loop for medical data synthesis. The pathologist can be replaced by training a reward model, which is used to fine-tune the generation model.

Strengths

- I admire that the authors broke down the barriers between disciplines, and I believe their proposed pathologist-in-the-loop synthetic data generation framework is a promising strategy for improving the generation of biomedical samples. It is the main reason why I give a positive score. As "pathologist-in-the-loop" is the key contribution, I personally encourage that the raw scores from doctors and the corresponding images should be published in pair after acceptance -- then we can check the quality of the manual annotations. - The paper is well-organized and the experiments are comprehensive.

Weaknesses

- The authors only gives the overall performances. I wonder if there is a direct verification for the reward model. - It would be better if this idea can be evaluated on other datasets. Of course, I know it is hard. - The technical innovation seems to be limited.

Questions

A direct evaluation of the reward model is helpful. I am curious about the relationship of the reward model quality and the synthesis performances.

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

The authors claimed that they have discussed the limitations, but it is not clear for me. Maybe I missed something.

Summary

This paper introduces a pathologist-in-the-loop framework for generating clinically plausible synthetic medical images. The training process is similar to generative adversarial nets, with two major modifications. First, the discriminator was trained by human input instead of real/fake labels. Second, the generator was replaced by Diffusion Models. Evaluation of synthetic bone marrow patches by expert hematopathologists, leveraging thousands of feedback points, demonstrates the significant quality enhancement achieved through human input.

Strengths

1. I found the paper to be well-written and informative, and I thoroughly enjoyed reading it. 2. The generation of synthetic data holds significant importance. 3. The extensive visualization of the synthetic images, along with the notable improvement resulting from the pathologist's involvement, is highly convincing.

Weaknesses

1. Missing an important baseline method that incorporates feedback from a real/fake binary classifier. This classifier can distinguish between plausible (real) and implausible (fake) images without relying on pathologists' feedback or cell-type labels (automated feedback). It is imperative to investigate whether utilizing this information can enhance the ability of Diffusion Models to generate realistic images. 2. The clinical plausibility criteria presented in Table 1 pose challenges for implementation in various other tasks. Firstly, developing the checklist items necessitates domain expertise. Secondly, disagreements among pathologists may arise regarding these criteria.

Questions

1. Despite incorporating 100% pathologist feedback, there appears to be a discernible disparity between synthetic and real data. Could you elaborate on the underlying reasons for this gap and propose potential solutions to mitigate it? 2. What if pathologists are unable to differentiate between real and synthetic images? For example, in some cases [1], the experts may not be able to distinguish the real/synthetic images. 3. What is the rate of implausible images after fine-tuning the diffusion model? 4. Consider integrating both real and synthetic data for training purposes. To what extent can this integration enhance the overall performance? This is particularly important as the authors have not yet achieved comparable performance of AI models trained solely on synthetic data when compared to those trained on real data. **Reference** [1] Hu, Qixin, Yixiong Chen, Junfei Xiao, Shuwen Sun, Jieneng Chen, Alan L. Yuille, and Zongwei Zhou. "Label-free liver tumor segmentation." In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 7422-7432. 2023.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

No potential negative societal impact was detected.

Summary

This paper proposes a pathologist-in-the-loop framework for generating clinically plausible synthetic medical images using diffusion models. The process involves pretraining a conditional diffusion model on real medical images, then using synthetic images labeled by expert pathologists to train a "clinical plausibility reward model" to predict pathologist feedback on new images. Finally, the diffusion model is fine-tuned using a reward-weighted likelihood loss that incorporates the reward model, to align synthetic outputs with clinical knowledge. The method is evaluated on a bone marrow cell image dataset. Results show the proposed method incorporating pathologist feedback significantly improves the clinical plausibility, fidelity, diversity, and downstream utility of synthetic images.

Strengths

1) Clinically plausible medical image synthesis is an interesting and important problem. This paper provides a promising human-in-the-loop framework to incorporate clinical knowledge. 2) The method is evaluated extensively both qualitatively and quantitatively. The results demonstrate clear improvements from incorporating expert feedback. 3) The paper is well-written and easy to follow. The method is described in sufficient detail.

Weaknesses

1) The criteria used by the pathologist for judging clinical plausibility could be described more precisely. Are there any quantitative metrics for each criterion? 2) Only binary feedback is collected from the pathologist. More fine-grained ratings could provide richer supervision signal. 3) Evaluations could be conducted on multiple datasets and with higher resolution images. 4) The comparison to "automatic feedback" using one classifier is somewhat weak. Better baselines or more comparisons could be evaluated.

Questions

1) The proposed method is only evaluated on a single dataset. Experiments on more datasets may be necessary to validate the generalizability of the method. Experimenting on higher resolution image synthesis could also be important for potential clinical usage. 2) It might be helpful to include more visual or qualitative comparisons between synthetic images marked as "implausible“ or "plausible". 3) Ye et al. also proposed a relevant idea using automatic feedback as the reward in [1]. It might be worth discussing this paper. [1] Ye et al. Synthetic Sample Selection via Reinforcement Learning. MICCAI 2020.

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

4 excellent

Contribution

3 good

Limitations

Authors could consider discussing the limitations and potential negative societal impact of this work: Limitations: 1) The method has only been evaluated on one medical imaging modality/dataset (bone marrow). Applicability to other datasets or modalities with different visual and clinical characteristics is unclear. 2) The framework requires extensive pathologist time for labeling synthetic images. Scalability to large diverse datasets may be problematic. 3) Binary plausibility labels may not capture nuanced aspects of clinical knowledge. More granular ratings could be beneficial. 4) The synthetic images are with low resolution (64x64). Quality and utility for diagnosis may degrade for higher resolution synthesis. Example negative societal impacts: 1) If not carefully validated, inaccuracies in synthetic medical images could mislead or harm practitioners relying on them for diagnosis/treatment. 2) Widespread availability of synthetic medical data could lead to privacy risks if reconstructed from patient data without consent.

After Rebuttal

Thank you for you reply, and the responses with revised paper solve my concerns. But, I still hope the author can refine the work mentioned in the second point and present the final results in the final version. Thank you!