xMIL: Insightful Explanations for Multiple Instance Learning in Histopathology

Summary

This work proposed an LRP-based extension for MIL setup in histopathology. It focuses on explanation quality, not derivation of novel architectures.

Strengths

It is quite well written in terms of language. Images are nicely depicting the key ideas.

Weaknesses

This paper does not follow the MIL paradigm. It references only standard MIL assumptions and methods dedicated to that. There is work extending standard MIL methods to other MIL assumptions such as additive MIL [1], presence and threshold-based MIL [2]. and lastly percentage MIL [3]. The novelty is extremely limited because I do not see here more than just the application of LRP to MIL models. Moreover, there are models such as CLAM [4] and ProtoMIL [5] that are addressing MIL in a more interpretable way, how are they working in your setup? Are they following the limitations? [1] Javed, Syed Ashar, et al. "Additive mil: Intrinsically interpretable multiple instance learning for pathology." Advances in Neural Information Processing Systems 35 (2022): 20689-20702. [2] Rymarczyk, Dawid, et al. "Kernel self-attention for weakly-supervised image classification using deep multiple instance learning." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2021. [3] Struski, Łukasz, et al. "ProMIL: Probabilistic multiple instance learning for medical imaging." ECAI 2023. IOS Press, 2023. 2210-2217. [4] Lu, Ming Y., et al. "Data-efficient and weakly supervised computational pathology on whole-slide images." Nature biomedical engineering 5.6 (2021): 555-570. [5] Rymarczyk, Dawid, et al. "Protomil: Multiple instance learning with prototypical parts for whole-slide image classification." Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Cham: Springer International Publishing, 2022.

Questions

At some point, you are relaxing MIL assumptions so much that is is not MIL anymore, especially when the bag is not a set of instances anymore, because the spatial dependence between instances is important. What is the difference between your work and a straightforward application of LRP to MIL problem?

Rating

5

Confidence

5

Soundness

1

Presentation

3

Contribution

1

Limitations

They are discussed.

Summary

The authors propose xMIL-LRP which is an explainable MIL method to overcome the shortcomings of previous instance score-based methods (such as attention-based). This can provide positive/negative evidence towards the label and can be slotted into existing MIL methods, since this is a post-hoc interpretability method.

Strengths

With most of MIL methods reaching saturation, especially with advent of transformer-based modules and very powerful patch-level foundation models, the performance gain to be had with new MIL method would likely be marginal. Therefore it is encouraging that the authors are proposing a tool for interpretability that can fit into any of these MIL methods (if I didn't misunderstand) - I am assuming this also works with any feature encoders. The LRP-based methods have been utilized many times outside this field, so I think it is positive that the authors brought it into the field with great rigor.

Weaknesses

There are few weakness/confusion that I would like the authors to address to make this a worthwhile contribution to the field. **Presentation**: I am assuming the authors are very familiar with LRP, but most of the audience in the computational pathology would not be. Therefore, I found lack of motivation/explanation beyone LRP quite confusing. I think Section 3.2 is too compressed (Especially lines 178-182). The section of faithful evaluation was also not easy to understand as this was entirely new analysis. How do these analyses relate to the fact that predicted evidence scores are positive/negative? - Can we do away with formalizing *evidence function*? It is a good formalization of the concept, but as soon as it was applied to the actual histology data, the concept was abandoned for the lack of ground truth, which I found it quite awkward. Why not go directly into introducing relevance scores (which do not need evidence function formalism) and allocating more explanation for general computational pathology readers? - I think explaining more rigorously how LRP would concretely be applied to AB-MIL and TransMIL (with few-step explanations) would be immensely helpful, since these are quite different architecture (multihead-attn vs. not) and authors simply bypass this by explaining the chain rules for "neurons". Only then readers can be convinced that LRP is easy to apply to any MIL of their choice. **Clarification vs. other methods**: My understanding is Additive MIL also provides positive/negative instance-level evidence (positive/negative logits) and can function very similarly as proposed LRP. I think the authors need to provide more explanation/experiments differentiating LRP from Additive MIL. Especially the heatmap ablations should probably include additive MIL results to convince the readers. In addition integrated gradient methods have been applied to explain positive/negative evidence towards bag-level labels [1], [2]. Can authors expand on this? **Restriction to binary classification task**: My understanding is that the current methodology is restricted to binary classification problem. For multi-class subtyping problem and survival prediction (where ordering of predicted risk is important), how can this be adapted? How would evidence scores add up to bag-level label, if the predicted endpoint is continuous (e.g. survival prediction)? I would like authors to expand on this, since only subset of the problems in pathology can be regarded as binary classification problem. **Speculation on different feature encoders**: While the explicit head-to-head comparison would not be necessary, I am curious to know how different feature encoders, especially with deluge of patch-level foundation models, affect LRP performance. Can authors share their thoughts on this? [1] Lee, Yongju, et al. "Derivation of prognostic contextual histopathological features from whole-slide images of tumours via graph deep learning." Nature Biomedical Engineering (2022): 1-15. [2] Song, Andrew H., et al. "Analysis of 3D pathology samples using weakly supervised AI." Cell 187.10 (2024): 2502-2520.

Questions

See the section on the weaknesses.

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

Please see above

Summary

The paper introduces xMIL-LRP, an efficient solution for explainable Multiple Instance Learning (MIL) by incorporating layer-wise relevance propagation (LRP) into MIL. The authors further employ the AH-rule to define the "propagation rules" for relevance. Experiments are conducted on three toy datasets and four real-world Whole Slide Image (WSI) prediction tasks, resulting in improved faithfulness scores for the explanations.

Strengths

1. The experimental setup on both toy datasets and real-world WSI tasks is robust, demonstrating the effectiveness of the proposed method. 2. The paper is logically structured, with well-illustrated figures that aid in understanding the results. 3. The development of explainable MIL methods is crucial for the deployment of these models in real-world applications, where understanding the model's decisions is important.

Weaknesses

Limited Novelty: The method primarily applies existing explainable AI techniques to the MIL domain, which limits its novelty. Unclear Method Description: Unclear Method Description: The definition of the relevance score $𝑝_{𝑖𝑗}$ as the attention score between neurons $j$ and $i$ is unclear. Attention-based MIL typically defines attention at the instance level, making it confusing how relevance is defined and propagated within this context. Absence of Fair Comparison with Additive MIL: The performance of additive MIL is directly cited without conducting new experiments under the same conditions (data splits, feature extractors, etc.). This could lead to an unfair comparison and underestimation of additive MIL's performance.

Questions

Please refer to weaknesses. One more question is that considering that xMIL-LRP potentially provides more explainable results, does it affect the prediction performance of the original MIL model? How do the authors balance between model explainability and prediction accuracy?

Rating

3

Confidence

3

Soundness

3

Presentation

2

Contribution

2

Limitations

N/A

Response

I appreciate the detailed responses provided, and I must admit that I had some misunderstandings when I was reviewing the paper. I now recognize that my argument about limited novelty no longer holds. However, I recommend that the authors incorporate a structured description of novelty into the paper, as it would enhance transparency and clarity. Excellent job on this aspect! Regarding the MIL formulation, I refer to a scenario where there is a bag of instances $ x = {x_i} $ with an associated known label $y $. For simplicity, let's assume $ y $ is binary, $ y \in {0,1}$. Each instance is independent and has its own hidden label $ y_i $. The bag is considered positive if at least one instance is labeled positive, which is the standard MIL assumption. Additionally, there are other MIL assumptions, such as: - Presence MIL: Multiple hidden labels must co-occur for the bag to be positive. - Threshold MIL: More than $ M $ instances within a bag must have a given hidden label. Both of these are well described in [1]. There is also a percentage MIL [2], among others. Crucial to the MIL definition is the independence of instances and their hidden labels constituting the bag-level label. From your manuscript, I understood that you work primarily with the Standard MIL assumption but relax the assumption about instance independence within a bag, suggesting that the order of instances within a bag is significant for xMIL. Was this a misunderstanding on my part based on your response? Regarding the other works I mentioned, I appreciate the comparison and acknowledgment provided. I would like to also point you to the application of MIL to the histopathology dataset that I've personally found interesting as the labels are derived based on the pathology change location on the tissue, as well as its severity - area covered by pathological change [3]. But this is just out of curiosity, no action is needed :) In general, you did well during the responses, I am keen to increase my grade, but first would like to see your comments regarding MIL formulations, and whether we agree on the definitions. References: [1] Rymarczyk, Dawid, et al. "Kernel self-attention for weakly-supervised image classification using deep multiple instance learning." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2021. [2] Struski, Łukasz, et al. "ProMIL: Probabilistic multiple instance learning for medical imaging." ECAI 2023. IOS Press, 2023. 2210-2217. [3] Rymarczyk, Dawid, et al. "Deep learning models capture histological disease activity in Crohn’s disease and ulcerative colitis with high fidelity." Journal of Crohn's and Colitis 18.4 (2024): 604-614.