- Rating
- 1
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
The authors investigate whether downstream models trained from radiology foundation models with a few annotated data samples learn biologically reasonable features. Specifically, they measure the prediction score change of samples with a lesion when the lesion region is occluded and compare it to prediction score changes when arbitrary regions are occluded. They find that in most cases, these prediction score changes/focus on correct regions do not increase despite increasing prediction task AUC. They further find that trained adapters (linear probe, attentive pooling, LoRA) often do not outperform training-free centroid-based baselines. Finally, they formulate an auxiliary loss based on the samples with occluded regions that leads to an improvement in the model's focus on the correct regions.
Critical strengths
- (S1) CCA attribution performance. The authors provide evidence that the CCA auxiliary loss seems to make the model focus on correct regions and improve model generalization performance (improved evidence peak inside the localization). This indicates that the method is actually useful for its intended purpose.
Critical weaknesses
- (W1) Effect of counterfactual intervention. The authors compute counterfactuals by overwriting a region of interest or sampled control region with a “local ring-mean fill”, which seems to result in a monotonous grey pixel area (see Figures 1, 3). I believe that this likely introduces unnatural image artefacts. Therefore, the counterfactual generation method may easily take an image out of the training distribution, which makes the representation of the image produced by the image encoder unpredictable and potentially unstable. Therefore, it is unclear if the resulting representation is still suitable for making a reasonable prediction. Similarly, any adapter trained from the unmodified medical images has not seen any of these grey patches at training time, so that an image with a grey patch at test time may result in unpredictable behavior. Overall, the resulting model prediction scores may not be interpretable. This renders any derived downstream measure (such as the evidence, the placebo scores, the win rates) potentially unreliable. To my understanding, the choice of the local ring-mean fill method is not discussed, and there is no evidence provided that the resulting images lead to reasonable representations/predictions, despite the availability of alternative occlusion methods and literature studying the effects of them (see, e.g, [1]). Without additional discussion and/or validation, I believe that there may not be sufficient evidence for the derived claims.
- (W2) Broad framing. The abstract and introduction frame the study as a general approach targeting medical foundation models. However, to my understanding, the main evidence is purely focused on binary detection tasks in radiology. How would this framework extend to, for example, multi-class classification, regression, time-to-event prediction, or other medical modalities beyond 2D chest radiography (and retinal fundus photographs)?
- (W3) Reliance on region-of-interest annotations. To my understanding, both the measuring framework and the CCA method rely on region-of-interest annotations for positive samples. For the latter, these are used as additional labels in the training process. In this sense, the improved external AUC and localization performances may also just be a result of using additional labels compared to the adapters that are trained purely on the classification labels. An additional baseline that uses the annotations as a direct supervision signal, if available, would be helpful to clarify the added benefit of the CCA loss formulation.
- (W4) Presentation. Personally, I struggled to understand many parts of the paper, as I found it difficult to read. The authors use plenty of previously unintroduced terminology that made it hard for me to follow the train of thought and to understand the details. For example:
- “Table 1 reports partial Spearman correlations against external AUC” -- This is the first mention of external AUC, and it is unclear to me how it is computed and why it is correlated to the introduced metric.
- “Calcification lesions carry higher edge energy than their paired controls (d = +0.80; Appendix D.5), so absolute win-rates are instrument-relative” -- What does “edge energy” mean? What “instrument” do the authors refer to in “instrument-relative”?
- “Because the scope condition lacks a priori predictability, CCA serves as an empirical test of whether the audited quantity is causally actionable” -- What is the “scope condition”? What is the “a priori predictability” in this context?
[1] Bluecher et al.: "Decoupling Pixel Flipping and Occlusion Strategy for Consistent XAI Benchmarks." TMLR, 2024.
Questions for authors
- (Q1) Outperforming the gate-0 method. The authors claim regarding the gate-0 method that “outperforming it is therefore a mathematical necessity, not a mere target. Failing to beat gate-0 implies the adapter extracts no novel features beyond first-order class statistics, regardless of its absolute AUC.” However, to my understanding, the features used to compute this gate-0 are deep representations from the feature encoders. Why are these considered “first-order class statistics”? A reason why an adapter may not outperform this baseline could be that the encoder produces strong representations that make positive and negative class samples easily separable in feature space. In this case, would this not be a strength of the encoder rather than a failure mode of the adapter?
- (Q2) Adapter evaluation. The authors mention that they compare “three adaptation depths (linear probe, attentive pooling, LoRA)”. How were the results presented in Table 1, Figure 2, Table 2, and Table 3 computed? Which adaptation method was used here? Or are they aggregated results across adapters? Or are those results from the gate-0 / gate-0b baselines?