Pith. sign in

REVIEW 4 major objections 4 minor 21 references

Composing VLM-selected clinical tools with a learned bottleneck model yields medical-image classifiers that are accurate, interpretable, and data-efficient.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A 'tool bottleneck' framework—VLM tool selection plus learned spatial fusion—matches or beats black-box classifiers, especially on scarce data.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection The Tool Bottleneck idea is real and the Camelyon17 results are credible, but the dermatology numbers lean on ground-truth annotations passed as 'tools,' and the paper's broad claims do not survive that. the 4 major comments →

arxiv 2512.21414 v2 pith:5HUVFJON submitted 2025-12-24 cs.CV cs.LG

A Tool Bottleneck Framework for Clinically-Informed and Interpretable Medical Image Understanding

Christina Liu , Alan Q. Wang , Joy Hsu , Jiajun Wu , Ehsan Adeli This is my paper
classification cs.CV cs.LG
keywords tool usevision-language modelsinterpretabilitymedical imagingconcept bottlenecktool compositiondata efficiencyhistopathology
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that medical image classifiers become more accurate and more interpretable when, instead of reasoning over raw pixels, a vision-language model selects clinically meaningful tools (e.g., nucleus detectors, lesion segmenters) and a small learned network fuses the tool outputs into a diagnosis. This Tool Bottleneck Framework replaces text-based tool composition, which struggles with spatially localized medical features, with a neural 'Tool Bottleneck Model.' On histopathology and dermatology tasks the framework matches or beats end-to-end CNNs and VLMs, and in data-limited regimes (4–64 labeled images) it shows the largest gains. The authors also show that the learned fusion makes tool importance measurable and decisions intervenable.

Core claim

The central claim is that composing a VLM-selected set of clinical tools through a learned bottleneck is a better inductive bias for medical image understanding than end-to-end learning or text-based tool composition. Concretely, the Tool Bottleneck Model rasterizes each tool's output into spatial maps, concatenates them, and feeds them to a CNN that outputs the prediction; tool knockout augmentation during training lets the model handle any VLM selection. Results on histopathology tumor detection and two dermatology classification tasks show accuracy/AUC on par or better than a standard CNN (e.g., 92.3% vs 88.6% on the histopathology task) and clear gains when only a few dozen labeled examp

What carries the argument

The Tool Bottleneck Model (TBM) is the central object: a frozen toolbox of clinically relevant tools produces feature maps, the TBM concatenates them across channels and applies a CNN to fuse them into an image-level prediction. Tool knockout augmentation—replacing unselected tools' outputs with a constant placeholder during training—makes the TBM robust to arbitrary tool selections and simultaneously learns the full conditional and every marginal conditional of the label given tool subsets.

Load-bearing premise

For the dermatology tasks, four of the seven tool channels are not produced by real tools but by ground-truth dermoscopic annotations treated as hypothetical detector outputs; if such detectors are unavailable or inaccurate in practice, the reported dermatology results are not established.

What would settle it

Run the dermatology experiments with the four dermoscopic structure maps produced by detectors trained on separate data (instead of ground-truth annotations) and compare AUC; if the framework's advantage over a standard CNN vanishes or reverses, the central claim fails for dermatology.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If TBF is correct, learned fusion of VLM-selected clinical tools is a viable alternative to end-to-end learning in medical imaging.
  • TBF's data efficiency means small clinical datasets (tens of images) can still yield usable classifiers if a relevant toolbox exists.
  • Tool-importance and intervention analyses give clinicians a way to inspect and correct model reasoning.
  • Text-based tool composition (code or natural language) is the limiting factor for medical tool-use frameworks.
  • Including all modality tools rather than VLM selection hurts performance, so tool selection matters.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The dermatology results rely in part on ground-truth structure annotations as hypothetical tool outputs; real deployment would require accurate detectors for pigment networks, streaks, etc., so the reported dermatology gains may reflect an upper bound.
  • The framework's value should extend beyond classification to other medical tasks with spatially localized features (e.g., segmentation, grading), if suitable tools exist.
  • A testable extension: use TBF with self-supervised or foundation-model tools to see whether the data-efficiency and interpretability benefits persist when tools are not domain-specific.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes the Tool Bottleneck Framework (TBF), which uses a medical VLM to select clinically relevant tools from a toolbox and then fuses the selected tools' outputs with a learned Tool Bottleneck Model (TBM) to make image-level predictions. TBF is evaluated on Camelyon17 (histopathology) and two ISIC 2017 binary tasks (dermatology), with reported accuracy/AUC on par with or better than CNNs and VLMs, plus data-efficiency and interpretability analyses. The method contribution is a learned composition mechanism for pixel-level and scalar tool outputs, enabled by a knockout-style training strategy that makes the TBM robust to arbitrary tool selections.

Significance. If the claims held, TBF would be a useful step toward clinically grounded and interpretable medical image classification: it generalizes Concept Bottleneck Models to spatially localized features, replaces text-based tool composition with a learned fusion network, and provides a simple strategy for handling arbitrary VLM tool selections. The authors are transparent in the appendix about the idealized nature of several dermatology tools, and the knockout proof in Appendix E is a sound adaptation of prior work. However, the headline claim—that TBF performs 'on par with or better than' black-box models and tool-use frameworks across histopathology and dermatology—is only cleanly supported on Camelyon17. Two of the three evaluation tasks rely on ground-truth dermoscopic structure annotations as 'hypothetical tool outputs' at test time, which is an oracle setup, not a realized tool-use system.

major comments (4)
  1. [Appendix C.2, Table 1, Sections 4.1 and 5.1] Four of the seven ISIC tools (pigment network, negative network, streaks, milia-like cysts) are not produced by learned detectors at test time; they are ground-truth dermoscopic feature annotations from the ISIC 2017 dataset. Appendix C.2 states this explicitly, calling it 'an idealized scenario in which accurate detectors ... 'exist''. These annotations are label-adjacent: milia-like cysts are strongly indicative of seborrheic keratosis (the non-melanocytic class in ISIC-MN), and pigment network is associated with melanocytic lesions. The TBM therefore receives structured label-adjacent information that the EfficientNet baseline does not. Since two of the three main tasks (ISIC-BM, ISIC-MN) and the data-efficiency experiment in Figure 2(b) depend on these oracle maps, the central claim of 'on par with or better than ... across histopathology and dermatology' is unsupported for a realize
  2. [Figure 2(b), Section 5.2] The data-efficiency argument is a key contribution claim, but the ISIC-MN panel is compromised by the same oracle-tool setup. The reported gains of TBF over EfficientNet in the 4–64 image regime may simply reflect that the TBM is fed ground-truth dermoscopic structure maps, whereas EfficientNet receives only raw pixels. These low-data results cannot be interpreted as evidence that the TBF framework, with realistically available tools, is more data-efficient. The claim 'TBF embodies clinically-relevant inductive biases' therefore needs to be re-established on Camelyon17 alone, or with predicted (non-oracle) ISIC tools. As written, the ISIC-MN data-efficiency result is an upper-bound experiment, not a demonstration of the framework.
  3. [Appendix D.1, Table 2 vs. Table 1] There is an internal inconsistency in the reported Camelyon17 accuracy for TBF with ImageNet pretraining: Table 1 reports 92.3%, while Table 2 reports 92.8%, and the text in D.1 says the accuracy 'increases to 92.3%'. The discrepancy needs to be resolved and the correct value reported consistently. If the difference is due to different evaluation splits or seeds, that should be stated.
  4. [Appendix B, Y-Net baseline] The paper repeatedly states that Y-Net is trained on 'the same amount and types of data' as TBF (Sections 4.3, 5.1, Appendix B). However, TBF receives five histopathology tool maps or seven dermatology tool maps, while Y-Net receives only a single segmentation mask (tumor masks for Camelyon, lesion masks for ISIC) plus raw images. This is not the same type or amount of pixel-level supervision. The comparison is still meaningful as a segmentation-for-classification baseline, but the 'same data' claim should be softened or the Y-Net training should be updated to consume the full tool-stack inputs.
minor comments (4)
  1. [Section 5.1, Table 1] The accuracy/AUC values are reported without confidence intervals or significance tests. Given the small differences in several comparisons (e.g., ISIC-MN 91.8 vs. 91.2), error bars or a significance statement would help the reader judge whether TBF is truly 'on par or better'.
  2. [Section 3.3, Equation (1)] The Bernoulli perturbation with parameter p = (1−α)0.5 + α s_i is described in words, but the notation is not fully formalized. Specifying the exact sampling distribution and how the placeholder value ¯z_i is applied would improve reproducibility.
  3. [Appendix C.2] The appendix uses the phrase 'hypothetical tool outputs' once, but the discussion of the four ground-truth structure maps would be clearer if the authors consistently labeled them as oracle inputs throughout the paper, including in the figure captions and table notes. Currently, Table 1 does not indicate that any ISIC results use ground-truth annotations.
  4. [Appendix D.1] The sentence 'TBF achieves 86.7% accuracy when trained from scratch, and increases to 92.3%' is inconsistent with Table 2, which lists 92.8% for the ImageNet-pretrained TBF. This is likely a typo, but it should be corrected.

Circularity Check

0 steps flagged

No equation-level circularity: the TBM training and knockout derivations are self-contained. The main caveat is an oracle-tool evaluation issue on ISIC, not a circular-derivation defect.

full rationale

The paper's derivation chain is not circular. The Tool Bottleneck Model is trained on frozen tool outputs: tools are fixed, the bottleneck tensor is the concatenation of tool maps, and f_theta is a CNN trained with cross-entropy or binary cross-entropy. The tool-knockout argument in Appendix E is an explicit adaptation of Nguyen et al. (2025), with the independence-of-mask derivation given in the paper itself; it does not presuppose the result. Cited prior work by the authors (Hsu et al., 2023, 2025) is contextual and not load-bearing; no uniqueness theorem or forced ansatz is imported from it. The central issue in the paper is the ISIC evaluation: Appendix C.2 states plainly that four of seven dermatology tool channels are 'ground-truth dermoscopic feature annotations provided by ISIC 2017 as hypothetical tool outputs—representing an idealized scenario in which accurate detectors for these clinically meaningful cues exist.' This means the reported ISIC-BM and ISIC-MN results, and the ISIC-MN data-efficiency curves, are an oracle upper bound rather than evidence about real tool use. That is a soundness and data-integrity limitation, not a circular derivation: the final label is not defined by the tool maps in the paper, and the model does not rename the tool outputs as predictions. The Camelyon17 experiments use predicted HoVer-Net outputs and do not share this caveat. Because the core TBM derivation is self-contained and the only serious concern is an oracle-evaluation caveat, the circularity score is low.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The framework rests on pretrained or oracle tools and a small number of hyperparameters (alpha, k, hand-set color thresholds). No new physical or conceptual entities are introduced beyond the TBM architecture itself.

free parameters (3)
  • alpha (Bernoulli perturbation strength) = 0.9 (Camelyon17, ISIC-BM), 0.8 (ISIC-MN)
    Controls the blend between the VLM tool-selection prior and random tool sampling during TBM training; tuned per task via grid search over {0.5,0.6,0.7,0.8,0.9} on validation sets.
  • k (number of VLM-selected tools) = 3 for all tasks
    Fixed number of tools selected by MedGemma per image; tuned via grid search over {2,3,4} on validation sets.
  • Handcrafted color-marker thresholds
    RGB thresholds and morphological area cutoffs in the derm color-marker tools are hand-set based on prior work, affecting two of the seven ISIC tool channels.
axioms (4)
  • domain assumption The VLM (MedGemma) produces useful tool selections for each image and task.
    Section 3.2/Appendix F: if the VLM selects irrelevant tools, the TBM must rely on the alpha perturbation and learned fusion to compensate.
  • domain assumption All tools in the toolbox are available as real detectors at test time and capture clinically relevant features.
    Section 3.1 states this; Appendix C.2 violates it for four ISIC tools by using ground-truth annotations instead of predicted tool outputs.
  • standard math Tool knockout augmentation lets f_theta learn all marginal conditionals p(Y | Z_{-M}).
    Appendix E adapts the Knockout result of Nguyen et al. (2025), relying on the mask M being independent of (Z, Y).
  • domain assumption HoVer-Net nucleus predictions transfer to Camelyon17 patches without fine-tuning.
    Appendix C.1 uses the frozen TIAToolbox HoVer-Net model; if domain shift is large, the tool feature maps are noisy or uninformative.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of A Tool Bottleneck Framework for Clinically-Informed and Interpretable Medical Image Understanding." pith.science (2026). https://pith.science/paper/5HUVFJON

@misc{pith2026251221414,
  author       = {Pith},
  title        = {Pith review of: A Tool Bottleneck Framework for Clinically-Informed and Interpretable Medical Image Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5HUVFJON}},
  note         = {Machine review of arXiv:2512.21414}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent tool-use frameworks powered by vision-language models (VLMs) improve image understanding by grounding model predictions with specialized tools. Broadly, these frameworks leverage VLMs and a pre-specified toolbox to decompose the prediction task into multiple tool calls (often deep learning models) which are composed to make a prediction. The dominant approach to composing tools is using text, via function calls embedded in VLM-generated code or natural language. However, these methods often perform poorly on medical image understanding, where salient information is encoded as spatially-localized features that are difficult to compose or fuse via text alone. To address this, we propose a tool-use framework for medical image understanding called the Tool Bottleneck Framework (TBF), which composes VLM-selected tools using a learned Tool Bottleneck Model (TBM). For a given image and task, TBF leverages an off-the-shelf medical VLM to select tools from a toolbox that each extract clinically-relevant features. Instead of text-based composition, these tools are composed by the TBM, which computes and fuses the tool outputs using a neural network before outputting the final prediction. We propose a simple and effective strategy for TBMs to make predictions with any arbitrary VLM tool selection. Overall, our framework not only improves tool-use in medical imaging contexts, but also yields more interpretable, clinically-grounded predictors. We evaluate TBF on tasks in histopathology and dermatology and find that these advantages enable our framework to perform on par with or better than deep learning-based classifiers, VLMs, and state-of-the-art tool-use frameworks, with particular gains in data-limited regimes. The project details and the code are available at https://christinaliu2020.github.io/tbm/.

Figures

Figures reproduced from arXiv: 2512.21414 by Alan Q. Wang, Christina Liu, Ehsan Adeli, Jiajun Wu, Joy Hsu.

Figure 1
Figure 1. Figure 1: Overview of our proposed Tool Bottleneck Framework. A VLM selects tools from a pre-specified toolbox of clinically-relevant tools. These tools are passed to a Tool Bottleneck Model, which composes/fuses the tool outputs to make a prediction. 1984). For any given task, two questions arise: (1) which features are most relevant, and (2) how should these features be integrated to make a prediction? Broadly, de… view at source ↗
Figure 2
Figure 2. Figure 2: Model performance of TBF vs. EfficientNet baseline over varying training set sizes in log scale. Mean ± 95% CI over seeds. TBF exhibits improved performance across all training set sizes. 5.3. Analysis We are interested in analyzing the “importance” of each tool for a given task and how that relates to the distribution of VLM tool selections during training. To measure importance of a given tool, we knocko… view at source ↗
Figure 3
Figure 3. Figure 3: Tool-wise importance (Eq. 1) and normalized frequency of VLM tool selections for TBM across Camelyon17 (left) and ISIC-BM/-MN (right). In each plot, the left axis shows the relative importance of each tool measured by the change in Accuracy (Camelyon17) and AUC (ISIC) when tools are individually removed during inference. The right axis shows the normalized frequency of tools selected by MedGemma during tra… view at source ↗
Figure 4
Figure 4. Figure 4: (a)-(e): Visualization of the nuclei–dropout intervention on two example Came￾lyon17 contour maps. For each example patch, we randomly remove individ￾ual nuclei by masking them out with probability pmask, which we sweep across pmask ∈ {0.0, 0.2, 0.4, 0.6, 0.8} (from left to right) to randomly mask out nuclei in the tool output maps. (f): As pmask is increased (dropout increased), the fraction of images wit… view at source ↗
Figure 5
Figure 5. Figure 5: Distribution of MedGemma selected tool combinations for Camelyon17 [PITH_FULL_IMAGE:figures/full_fig_p023_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Distribution of MedGemma selected tools combinations for top 3 tools across the ISIC 2017 tasks that M is sampled independently of (Z, Y ). ZM denotes Z with elements knocked-out according to indices in M, and Z−M denotes all non-knocked-out elements. We denote the full conditional as p(Y | Z) and all marginal conditionals as p(Y | Z−M). During training, we construct knockout-augmented inputs Z ′ (M, Z) = … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 8 linked inside Pith

  1. [1]

    h i s t o _ n u c _ c e n t r o i d

    As mentioned in the main paper, this is easily implemented by replacing the tool outputzi with¯zi of the same shape. 24 Tool Bottleneck Framework To see thatf θ learns all such marginals simultaneously, consider the expected training loss under tool knockout: L(θ) =E Z′,Y EM ℓ Y, fθ(Z′(M,Z) (4) =E Z,Y EM X m∈M I(M=m)ℓ Y, fθ(Z′(m,Z)) (5) =E Z,Y X m∈M p(M=m...

    2017

  2. [3]

    21 Liu W ang Hsu Wu Adeli Table 3: Performance of TBF ablations across Camelyon17 (Accuracy) and ISIC 2017 (AUC). Model Camelyon17 ISIC-BM ISIC-MN Bernoulli 90.3 72.7 90.1 α= 0.5 90.5 75.3 88.9 α= 0.6 91.3 75.1 89.1 α= 0.7 90.7 75.6 91.2 α= 0.8 91.6 75.3 91.9 α= 0.992.3 77.090.6 α= 1.0 90.2 74.9 90.4 Random top-k91.7 76.192.0 Dynamic top-k91.4 66.5 90.4 A...

    2017

  3. [4]

    missing” value of−1 (a constant map of −1s) and are handled consistently across all masking regimes. Details on filling dropped tools with the “missing

    As described in Section 4.1, all Camelyon17 tool maps are scaled to lie in [0,1] before masking; dropped tools are represented by channels filled with a constant “missing” value of−1 (a constant map of −1s) and are handled consistently across all masking regimes. Details on filling dropped tools with the “missing” values are provided in Section E. C.2. IS...

    2017

  4. [6]

    Appendix E. Details on Tool Knockout We show that our tool knockout augmentation enablesf θ to estimate the full conditional (the distribution ofY=yconditioned on all tool outputs) and all marginal conditionals (the distributionY=yconditioned on any subset of tool outputs). Our argument follows the theoretical analysis of Knockout by Nguyen et al. (Nguyen...

    2025

  5. [9]

    Correlation of nuclear morphometry with progres- sion of breast cancer.Cancer, 68(9):2012–2016,

    Kenneth J Pienta and Donald S Coffey. Correlation of nuclear morphometry with progres- sion of breast cancer.Cancer, 68(9):2012–2016,

    2012

  6. [10]

    Medgemma technical report.arXiv preprint arXiv:2507.05201,

    Andrew Sellergren, Sahar Kazemzadeh, Tiam Jaroensri, Atilla Kiraly, Madeleine Traverse, Timo Kohlberger, Shawn Xu, Fayaz Jamil, C ´ ıan Hughes, Charles Lau, et al. Medgemma technical report.arXiv preprint arXiv:2507.05201,

    Pith/arXiv arXiv

  7. [11]

    Medclip: Contrastive learning from unpaired medical images and text

    Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. InProceedings of the Conference on Empirical Methods in Natural Language Processing. Conference on Empirical Methods in Natural Language Processing, volume 2022, page 3876,

    2022

  8. [13]

    Robust and interpretable medical image classifiers via concept bottleneck models.arXiv preprint arXiv:2310.03182,

    An Yan, Yu Wang, Yiwu Zhong, Zexue He, Petros Karypis, Zihan Wang, Chengyu Dong, Amilcare Gentili, Chun-Nan Hsu, Jingbo Shang, et al. Robust and interpretable medical image classifiers via concept bottleneck models.arXiv preprint arXiv:2310.03182,

    Pith/arXiv arXiv

  9. [14]

    Each WSI is manually annotated by pathologists to mark tumor regions, from which non-overlapping 96×96 pixel patches are extracted and labeled as eitherTumororNormal

    is adapted from the CAME- LYON17 challenge (Litjens et al., 2018), which consists of whole-slide images (WSIs) of breast cancer metastases in lymph node sections. Each WSI is manually annotated by pathologists to mark tumor regions, from which non-overlapping 96×96 pixel patches are extracted and labeled as eitherTumororNormal. A patch is labeledTu- morif...

    2018

  10. [16]

    tumor”, “no tumor

    for 25 epochs, and monitored with Dice score. In Stage 2, we jointly train the segmentation and classification head with the combined lossL=L seg +L cls,where bothL seg andL cls are cross-entropy losses for the segmentation mask and binary label, respectively. Stage 2 uses Adam with a lower learning rate (3×10 −4), batch size 16, and 20 epochs. For both d...

    2023

  11. [19]

    22 Tool Bottleneck Framework Figure 5: Distribution of MedGemma selected tool combinations for Camelyon17 D.4. T ool Output Intervention Besides tool importance discussed in Section 5.3, another method for interrogating TBF’s decision-making is to intervene or manipulate the tool outputs, since the features encoded in the tool outputs are clinically meani...

    2013

  12. [1984]

    Llava-med: Training a large language- and-vision assistant for biomedicine in one day.arXiv preprint arXiv:2306.00890,

    12 Tool Bottleneck Framework Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tris- tan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language- and-vision assistant for biomedicine in one day.arXiv preprint arXiv:2306.00890,

    Pith/arXiv arXiv

  13. [2003]

    Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 interna- tional symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isi...

    2017

  14. [2012]

    shades of gray

    and color-constancy-based preprocessing for robust skin image analysis (Ciurea and Funt, 2003). Following these approaches, we first apply a simple “shades of gray” color con- stancy transform (Ciurea and Funt,

    2003

  15. [2016]

    Mobilenets: Efficient convolutional neural networks for mobile vision applications.arXiv preprint arXiv:1704.04861,

    Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, To- bias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications.arXiv preprint arXiv:1704.04861,

    Pith/arXiv arXiv

  16. [2017]

    Probabilistic concept bottleneck models.arXiv preprint arXiv:2306.01574, 2023a

    Eunji Kim, Dahuin Jung, Sangha Park, Siwon Kim, and Sungroh Yoon. Probabilistic concept bottleneck models.arXiv preprint arXiv:2306.01574, 2023a. Injae Kim, Jongha Kim, Joonmyung Choi, and Hyunwoo J. Kim. Concept bottleneck with visual concept filtering for explainable medical image classification.arXiv preprint arXiv:2308.11920, 2023b. Accepted to MedAGI...

    Pith/arXiv arXiv

  17. [2018]

    X Liu, J Wu, X Luo, W Liao, S Zhang, and G Wang

    ISSN 2047-217X. X Liu, J Wu, X Luo, W Liao, S Zhang, and G Wang. 2024 ieee international symposium on biomedical imaging (isbi).IEEE International Symposium on Biomedical Imaging (ISBI),

    2047

  18. [2021]

    Rsvp: Reasoning segmentation via visual prompting and multi-modal chain-of-thought.arXiv preprint arXiv:2506.04277,

    Yi Lu, Jiawang Cao, Yongliang Wu, Bozheng Li, Licheng Tang, Yangguang Ji, Chong Wu, Jay Wu, and Wenbo Zhu. Rsvp: Reasoning segmentation via visual prompting and multi-modal chain-of-thought.arXiv preprint arXiv:2506.04277,

    Pith/arXiv arXiv

  19. [2022]

    Graph concept bottleneck models.arXiv preprint arXiv:2508.14255,

    Haotian Xu, Tsui-Wei Weng, Lam M Nguyen, and Tengfei Ma. Graph concept bottleneck models.arXiv preprint arXiv:2508.14255,

    Pith/arXiv arXiv

  20. [2023]

    Integrating Clinical Knowl- edge into Concept Bottleneck Models

    13 Liu W ang Hsu Wu Adeli Winnie Pang, Xueyi Ke, Satoshi Tsutsui, and Bihan Wen. Integrating Clinical Knowl- edge into Concept Bottleneck Models. InProceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, volume LNCS 15004, pages 243–253. Springer Nature Switzerland, October

    2024

  21. [2025]

    Label-free concept bottleneck models.arXiv preprint arXiv:2304.06129,

    Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models.arXiv preprint arXiv:2304.06129,

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.