MAPLE: Multi-scale Attribute-enhanced Prompt Learning for Few-shot Whole Slide Image Classification

We thank the reviewer for the comments and for the time spent reviewing our paper. We address the weaknesses (W) and questions (Q) as follows:

---

[W1] Clarification on Figure 1.

We thank the reviewer for pointing this out. Actually, Figure 1 is intended to emphasize the key difference in alignment strategy between our MAPLE and the existing studies (e.g., TOP and ViLa-MIL), where the existing methods only perform the slide-level alignment (shown in Fig.1(a)), while our method considers additional entity-level features and incorporates subtype-specific phenotypic attributes for more interpretable and precise alignment (shown in Fig.1(b)). Since multi-scale modeling also appears in methods like ViLa-MIL, we do not explicitly distinguish this in the figure, but note it in the caption. We will further revise the figure and related text in the final version to avoid confusion.

[W2] Concerns on texts in Section 2.2.

We acknowledge that different methods may adopt varying strategies to compute logits, often involving additional architectural components and complex aggregation strategies. However, the purpose of Section 2.2 is to provide a concise and coherent overview of the general paradigm, rather than to detail the full implementation of all existing approaches. Therefore, we believe that our formulation captures the essential inference process commonly adopted in prompt-based few-shot WSI classification. We would be happy to further clarify any part of this formulation if needed.

[Q1] Reliability of LLM and insights of choosing appropriate LLMs.

We acknowledge the reviewer's concern regarding the reliability of LLM-generated prompts, and we provide the detailed discussion in the response to [W3] of Reviewer dMRb. To assess the reliability of our generated prompts, we examine the entities generated from GPT4 by querying different LLMs (e.g., Claude 3.5 Sonnet, DeepSeek-V3 and Qwen2.5) using the prompt: "Do you think [ENTITY] is a key entity relevant to distinguishing different subtypes in [DATASET]?" All of these LLMs affirm the relevance and importance of the generated entities, which can demonstrate the accuracy of selected entities.

For practical usage of choosing appropriate LLMs in clinical scenarios, we suggest selecting recently released LLMs that benefit from improved medical knowledge coverage and instruction-following ability. In addition, combining different outputs from multiple LLMs and incorporating domain expert verification may offer a more robust and trustworthy solution than relying solely on a single model. We consider a more systematic exploration of the reliability and choice of different LLMs is an important direction for future work.

[W3 and Q2] The choice of using PLIP and CONCH as VLMs.

Vision-language models (VLMs) have experienced explosive growth in the visual domain, however, CLIP is still widely regarded as a standard backbone for few-shot learning research [1][2]. In this work, we further explore the applicability of CLIP and its variant in pathology domain, PLIP, as our backbone. PLIP retains the architecture of CLIP while being fine-tuned on large-scale pathology-specific image-text data, making it suitable for investigating the effectiveness of prompt learning in few-shot WSI classification. Compared to CONCH, which is based on the more complex CoCa architecture, PLIP has lighter parameters and a structure more consistent with CLIP, facilitating direct and fair comparisons.

We appreciate the reviewer’s suggestion to consider other strong pathology foundation models such as UNI, GPFM, and Virchow. However, we note that these are image-only models that do not include a compatible text encoder, and therefore cannot be directly integrated into our prompt-learning pipeline, which relies on vision–language alignment. To this end, we instead conduct the ablation study using pathology VLMs, including CLIP-based methods (QuiltNet, PLIP) and the CoCa-based method (CONCH). As shown in the table below, CONCH provides the best classification performance as it is a stronger foundation model.

Next, for the purpose of illustrating the advantage of our MAPLE under different foundation models, we also compare it with all baselines using CONCH as the backbone. The experimental results shown in the response to [Q4] of Reviewer TGVt clearly verify the advantage of our MAPLE for few-shot WSI classification. We will include these CONCH-based results in the revised paper.

In summary, we can observe that the choice of different VLM backbones could affect the few-shot WSI classification results. However, our MAPLE can achieve consistently superior prediction results under different vision-language foundation models (e.g., CLIP, PLIP, CONCH), highlighting the advantages of our MAPLE to jointly integrate multi-scale visual semantics and perform prediction at both the entity and slide levels.

[1] Zeng et al. Local-Prompt: Extensible Local Prompts for Few-Shot Out-of-Distribution Detection. ICLR 2025.
[2] Pan et al. NLPrompt: Noise-Label Prompt Learning for Vision-Language Models. CVPR 2025.

[Q3] Results in Table 1 are relatively low, which stems from the choice of feature extractor.

We appreciate the reviewer’s observation. We agree that CONCH is a significantly stronger foundation model compared to PLIP, which may partly explain the AUC difference observed in Table 1 of original paper. To address this, we conduct additional experiments using CONCH as the visual encoder, with results presented in the response to [Q4] of Reviewer TGVt. Notably, MAPLE still achieves the best performance under this stronger backbone, reinforcing the effectiveness and generalizability of our method across different VLMs. We will include these CONCH-based results in the revised paper.

[Q4] Explanation on visualization details in Section 4.3.

The visualized patches are selected based on attention scores from the entity-guided cross-attention module. For each entity, we rank the attention scores and visualize the top-k scoring patches, which reflect the most relevant regions associated with that entity. We will clarify this detail explicitly in Section 4.3 of the revised paper.

[Q5] The point of using the global description and WSI-level representation if the entity can be precisely identified.

In MAPLE, entity-level prompts are designed to capture fine-grained and localized entities. However, WSIs are inherently heterogeneous, and accurate diagnosis often requires integrating evidence from multiple spatial and semantic cues, rather than relying on isolated entities alone. For instance, the presence of necrosis alone may not be diagnostically conclusive. However, when necrosis co-occurs with abnormal mitotic activity and specific inflammatory patterns, the combination strongly supports malignancy or tumor subtyping [1]. In such cases, global representations are essential for capturing diagnostic dependencies and interrelationships across the entire slide. Based on the above consideration, our slide-level representation is designed specifically to provide a holistic view that aggregates entity features. As confirmed in Table 2 of the main paper, MAPLE combined with entity-level and slide-level representations consistently outperforms using either alone, highlighting the effectiveness of both entity-level and slide-level information.

[1] Wesseling et al. The pathological diagnosis of diffuse gliomas: towards a smart synthesis of microscopic and molecular information in a multidisciplinary context[J]. Diagnostic Histopathology, 2011, 17(11): 486-494.

We thank the reviewer for the comments and for the time spent reviewing our paper. We address the weaknesses (W) and questions (Q) as follows:

---

[W1 and Q1] Prompt reliability and MAPLE’s hierarchy.

We appreciate the reviewer’s insightful comments regarding the reliability of LLM-generated prompts and the clinical plausibility of MAPLE’s hierarchical design.

Regarding the first point, we address this concern in detail in our response to [W3] of Reviewer dMRb. In current step, we carefully design the LLM instructions in the Prompt Construction stage and evaluate the reliability of GPT-4 outputs by querying multiple independent LLMs (e.g., Claude 3.5 Sonnet, DeepSeek-V3, and Qwen2.5), aiming to minimize the potential impact of hallucinations and increase the trustworthiness of the extracted entities. In future work, we plan to incorporate expert correction and clinical guidelines to further refine and validate entity construction.

On the second point, MAPLE’s hierarchical design is motivated by real-world diagnostic workflows. We clarify this hierarchy and provide more explanations to validate it. Specifically, the hierarchy in MAPLE is reflected in two aspects: multi-scale (high magnification and low resolution) and multi-level (entity-level and slide-level) design:

• One the one hand, at high magnification, pathologists analyze cellular components such as nuclear pleomorphism and cytoplasmic features for cancer diagnosis. As to the image at low magnification, they distinguish different cancer subtypes by examining tissue architecture such as gland formation and tumor-stroma interfaces [1][2].
• On the other hand, clinical diagnosis is rarely based on isolated observations. Instead, pathologists synthesize cues from multiple key tissue entities (e.g., mitotic activity, necrosis, glandular structures) and integrate them with global tissue context to reach a diagnosis [3]. For instance, necrosis alone may be inconclusive, but when co-occurring with abnormal mitoses and specific inflammatory patterns, the combination supports malignancy or tumor subtyping [4]. Our design emulates this diagnostic reasoning by jointly modeling localized entity-level and global slide-level descriptions.

[1] Kumar et al. Whole slide imaging (wsi) in pathology: current perspectives and future directions. Journal of digital imaging, 33(4):1034–1040, 2020.
[2] Li et al. A multi-resolution model for histopathology image classification and localization with multiple instance learning. Computers in biology and medicine, 131:104253, 2021.
[3] Heba et al. A comprehensive review of the deep learning-based tumor analysis approaches in histopathological images: segmentation, classification and multi-learning tasks. Cluster Computing 2023.
[4] Wesseling et al. The pathological diagnosis of diffuse gliomas: towards a smart synthesis of microscopic and molecular information in a multidisciplinary context[J]. Diagnostic Histopathology, 2011, 17(11): 486-494.

[W2 and Q2] Lack direct quantification on how well selected patches match textual prompts.

To the best of our knowledge, public WSI datasets typically lack spatial annotations that directly correspond to textual entities, making it infeasible to compute IoU or Dice metrics. In our current setup, we follow common practice adopted by prior works [1, 2], where semantic alignment is indirectly validated by visualizing the top-k patches with the highest attention scores from the entity-guided module and confirming that they visually reflect the corresponding textual descriptions.

To further address the reviewer’s concern and provide quantitative insights, we additionally compute the similarity scores between selected patches and their corresponding entity descriptions. Specifically, for each entity, we select the top-50 and bottom-50 patches based on the similarity scores from entity-guided attention, and then use the CONCH model to compute similarity scores with the entity description and the patches from different sets. Finally, we can derive the average scores of the top-50 and bottom-50 patches. We report the scores for the examples (Figure 4 of the original paper) in the table below. The results confirm that the selected patches align well with their associated textual descriptions, further validating the effectiveness of our method.

[1] Lu et al. Visual Language Pretrained Multiple Instance Zero-Shot Transfer for Histopathology Images. CVPR 2023.
[2] Jaume et al. Modeling Dense Multimodal Interactions Between Biological Pathways and Histology for Survival Prediction. CVPR 2024.

[Q3] Complementarity of multi-scale features and selection of magnifications.

MAPLE is designed to capture distinct histological entities at different magnifications. Different resolution levels in WSIs naturally correspond to different semantic scales: low magnification (e.g., 5×) reveals tissue architecture and global organizational patterns, while high magnification (e.g., 10×) focuses on cellular morphology and nuclear detail. These features are inherently complementary rather than redundant. As demonstrated in Table 2 of the original paper, integrating both scales consistently outperforms using either scale alone. To further validate this observation, we perform the paired t-test comparing multi-scale and single-scale results. The resulting p-values are consistently below 0.05, confirming that the performance gains from multi-scale integration are statistically significant.

Our resolution choices follow ViLa-MIL [1], which also employs dual-scale inputs at 5× and 10× for few-shot slide-level classification. We agree that higher magnification (e.g., 20×) may offer finer morphological cues. To explore this, we conduct experiments using 20× as the high resolution, and find that 10× and 20× perform comparably, suggesting that both magnifications can effectively capture high-resolution histological features. Considering that 20× images significantly increase memory usage and computational cost, we adopt 10× in our main experiments to achieve a balance between performance and efficiency.

[1] Shi et al. Vila-mil: Dual-scale vision-language multiple instance learning for whole slide image classification. CVPR 2024.

[W3 and Q4] Comparison against existing multi-scale MIL methods.

We appreciate the reviewer’s suggestion and conduct additional experiments to include existing multi-scale MIL baselines such as DTFD-MIL, Dual-Stream MIL, and Cross-Scale MIL under the same 5×/10× magnification settings used in MAPLE. As shown in the table below, MAPLE consistently outperforms these baselines across all datasets, highlighting the advantage of our multi-scale prompt-guided VLM-based method. These results further confirm that the performance gain of MAPLE arises not merely from multi-scale integration, but from the combination of multi-scale modeling and language-guide prompt supervision via vision-language models.

We thank the reviewer for the comments and for the time spent reviewing our paper. We address the weaknesses (W) as follows:

---

[W1] The font size in Figure 1 is too small and difficult to read.

We thank the reviewer for pointing this out. We will revise Figure 1 by enlarging the font size and optimizing the layout to ensure clarity and readability in the final manuscript.

[W2] In Figure 2, the meaning of the arrows between the entity-level features and the learnable prompts is unclear.

We appreciate the reviewer’s suggestion. The arrows represent the computation of similarity scores between the extracted entity-level features and the learnable prompts. We will revise Figure 2 by adding appropriate annotations to clarify their meaning and ensure interpretability.

[W3] Reliability and completeness of entity extraction by LLMs.

We thank the reviewer for this insightful comment. To ensure comprehensive and accurate extraction of entities, we carefully design the LLM instructions in the Prompt Construction stage to prioritize entities with higher clinical importance. For example, we incorporate prompts such as "iteratively suggest a discriminative histological entity not in the current entity set" (see Appendix B.1 for details), thereby reducing the likelihood of missing critical entities. This prioritization encourages the LLM to focus on highly discriminative entities.

To further evaluate the accuracy of the selected entities, we examine the entities generated from GPT4 by querying different LLMs (e.g., Claude 3.5 Sonnet, DeepSeek-V3 and Qwen2.5) using the prompt: "Do you think [ENTITY] is a key entity relevant to distinguishing different subtypes in [DATASET]?" All of these LLMs affirm the relevance and importance of the generated entities, which can demonstrate the accuracy of selected entities.

Furthermore, as shown in our ablation study on the number of entities (Figure 8 in the original paper), increasing the number of entities beyond a certain threshold does not improve performance and may even slightly decrease. This suggests that redundant entities may lack discriminative phenotypic attributes relevant to cancer subtype classification and potentially introduce noise that hinders the model's decision-making process. This observation supports the completeness and effectiveness of our selected entity set.

As acknowledged in our Limitations section, we recognize that LLM-generated entities may still suffer from hallucinations or generate clinically irrelevant descriptions in the absence of expert curation. To mitigate this, we propose the following future directions:

• Incorporating expert review or clinical guidelines to refine and validate the extracted entities;
• Aggregating outputs from multiple LLMs (e.g., a mixture-of-experts approach) to achieve more robust and reliable entity sets and descriptions rather than relying solely on a single model.

We believe these strategies will enhance the reliability and clinical applicability of LLM-driven entity construction, and we plan to explore them in future versions of MAPLE.

[W4] Patch overlap and the region selection module.

The Region Selection module in MAPLE is designed to accurately identify tumor-relevant regions while suppressing irrelevant background areas, thereby guiding the model to focus on informative patches that align well with the entity-level prompts. Since patch instances are partitioned in a non-overlapping manner and the region selection ratio is set to r = 0.7, the selected patches can effectively cover a variety of heterogeneous tumor-related regions within a WSI, and thus do not exhibit a high degree of spatial redundancy or overlap. We also investigate the impact of r in our ablation study (Section E.3 of the main paper). As shown in Figure 9, higher values of r (e.g., r > 0.7) yields the slight performance degradation. This decline can be attributed to the inclusion of less informative or non-tumor regions that compromise the quality of the selected patch set. In contrast, lower values (e.g., r < 0.7) tend to overly concentrate on a few localized regions, resulting in performance degradation.

We thank the reviewer for the comments and for the time spent reviewing our paper. We address the weaknesses (W) and questions (Q) as follows:

---

[W1 and Q1] SOTA pathology foundation models for MIL baselines.

We agree that traditional MIL approaches like ABMIL and TransMIL could indeed benefit from stronger image-only foundation models such as Virchow 2 or UNI2-h. In our main experiments (Table 1 and 5), we use the same feature extractor (i.e., PLIP and CLIP) across all baselines and MAPLE to ensure a fair comparison. To address the reviewer's concern, we additionally conduct the experiment including ABMIL and TransMIL with stronger image-only foundation models (e.g., UNI2-h), and MAPLE with a stronger VLM (e.g., CONCH). As shown in the table below, we can derive the following observations. Firstly, both ABMIL and TransMIL with UNI2-h exhibit noticeable performance gains, validating the importance of encoder strength. Secondly, MAPLE with CONCH can consistently beat the MIL based methods with latest image-only foundation model (i.e., UNI2-h), demonstrating the scalability and effectiveness of our method. Moreover, we compare MAPLE with all baseline methods under CONCH for a fair comparison in the response to [Q4]. MAPLE achieves the best performance across different few-shot settings and datasets, highlighting its effectiveness under different VLMs. In summary, these findings confirm that while backbone strength contributes to overall performance, MAPLE consistently delivers superior results.

[Q2 and W2] Generalizability beyond 16-shot settings.

For few-shot WSI classification, we follow prior works (e.g., MSCPT and FOCUS) and set the shots as 4, 8 and 16. We agree that evaluating the performance under larger-shot settings (e.g., 32-shot and 64-shot) is helpful for assessing MAPLE to more realistic scenarios where more WSIs per class are available. To this end, we conduct additional experiments under 32-shot and 64-shot settings, and present the results in the table below. As shown, MAPLE continues to outperform existing baselines, demonstrating its generalizability under larger-shot settings. Due to character limitation, we report the fully supervised results directly in the text. Specifically, our MAPLE achieves the best AUC of 0.903, 0.990 and 0.966 across three datasets, demonstrating significant improvements over the second-best method FOCUS with AUC of 0.880, 0.975 and 0.955.

[W3 and Q3] Comparison with ConcepPath.

We thank the reviewer for pointing out the relevance of ConcepPath. We report the results of ConcepPath in the table below. MAPLE consistently outperforms ConcepPath across different datasets and few-shot settings, highlighting the effectiveness of our method in few-shot WSI classification.

[Q4] Evaluating the effect of using different VLMs.

In addition to the experiments using CLIP and PLIP in the original paper, we provide further results with stronger pathology VLMs such as CONCH in the table below. MAPLE consistently outperforms baseline methods across different datasets and few-shot settings, highlighting the effectiveness of our method. We will include these results in the revised paper. Furthermore, we conduct a comprehensive ablation study evaluating the impact of different pathology VLMs including QuiltNet, PLIP, and CONCH, as detailed in our response to [W3 and Q2] of Reviewer yzxQ.