GRAPE: Graph-Augmented Prototype Explanations for Interactive Medical Image Diagnosis

arXiv:2606.30901 · cs.CV · Submitted 2026-06-29 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "GRAPE: Graph-Augmented Prototype Explanations for Interactive Medical Image Diagnosis".

Jane: Here's the gist of the paper. Prototype-based medical image classifiers present three clinical limitations: they treat findings as independent, silently amplify unsafe physician feedback, and require full retraining whenever a new finding is needed The authors present GRAPE (Graph-Augmented Prototype Explanations), a unified architecture that addresses all three challenges.

Tom: Addressing Clinical Limitations The limitations addressed by GRAPE include independent concepts, unsafe spatial interaction, and closed concept vocabulary.

Jane: The proposed solution involves a Graph Attention Task Head to model anatomical concept co-occurrence a Concept-Mismatch Safety Check to warn when the model’s dominant finding inside a doctor-drawn region conflicts with the claimed label and Open-Vocabulary Prototype Anchoring to align visual prototypes to clinical text.

Tom: How the Training Signal is Built The GRAPE training framework consists of a four-stage pipeline.

Title and authors: Tom: So we're talking about this new paper, "GRAPE: Graph-Augmented Prototype Explanations for Interactive Medical Image Diagnosis," and it tackles three major problems with how we currently use prototype-based medical image classifiers.

Jane: It sounds like they are fixing issues where the system treats findings as totally independent, where it lets unsafe physician feedback silently amplify errors, and where adding a new finding requires completely retraining the entire atlas from scratch.

Lu: That's right, and this paper proposes GRAPE as one unified architecture to solve all three of those things at once.

Meng: So what does this mean for the actual deployment side? Does it just make the model more accurate, or is there a bigger structural fix happening here?

Tom: Well, they introduce a Graph Attention Task Head which models how anatomical concepts actually co-occur, and they have this Concept-Mismatch Safety Check to warn doctors when what the model thinks is dominant inside a drawn area doesn't match what the doctor labeled it as.

Jane: That safety check sounds really interesting because it directly addresses that problem of unsafe feedback you mentioned earlier by catching errors before they get cemented into the system.

Lu: Exactly, and on top of that, they have Open-Vocabulary Prototype Anchoring which lets them align visual prototypes to clinical text descriptions, so you can add a new finding without needing to retrain everything.

Meng: That zero-shot addition capability is huge for production; it means we don't have to freeze the whole system just because a doctor sees something new they want the model to recognize.

Tom: And on the performance side, this Graph Attention Task Head actually boosted macro F1 by plus thirteen point eight percentage points over their prototype baseline on a dataset called TBX11K.

Jane: Thirteen point eight is a big jump, and that's because they are using the graph to understand the relationships between concepts like how active tuberculosis and cavity formation co-occur.

Lu: They build this data-driven co-occurrence graph from training labels, which they call the K times K conditional probability matrix A prime equals P of k given k prime, which is thresholded at zero point one to get a directed binary adjacency with self loops where each node’s incoming edge weights sum to one.

Meng: I need to know if building that graph adds significant computational overhead during inference, because running things at scale is a big deal for me.

Title and authors: Tom: They actually found that the Graph Attention Task Head adds only one millisecond over the linear head when you're running at a batch size of one, which keeps it pretty real-time friendly.

Jane: That’s good news for interactive diagnosis, because if it’s fast enough to interact with a doctor in real time, that makes the whole thing much more useful than just having high accuracy on a static test set.

Lu: They also have this four stage training framework where they start with concept supervision, then extract concept vectors using CAM-weighted spatial averages of frozen backbone features, and then focus on prototype learning with a loss L three which includes both the contrastive prototype learning loss and the VLM alignment loss.

Meng: So the training process is modular, starting from basic supervision and moving into these more complex graph structures before finally hitting that prototype optimization step.

Tom: The most impressive part for me, looking at their results on TBX11K and NIH ChestX-ray14, is that the Concept Mismatch Safety Check caught eighty-five percent of erroneous annotations compared to fifty-one percent for MC Dropout with zero extra inference cost.

Jane: Eighty-five percent is a big win there because it shows how much safer the system becomes when you let doctors correct its initial guesses.

Lu: And they also showed that using the Uk gated variant for the safety check resulted in a false positive rate of sixteen point four percent, which they found to be quite low.

Meng: So if we're talking about deploying this, I'm worried about the cost of maintaining that graph structure; how much memory does that co-occurrence graph take up?

Tom: They focused on making it efficient by using the Graph Attention Task Head to build that relationship map, and they showed the GAT edge importance ablation where the self loop on concept zero, which is active tuberculosis, was identified as the most informative edge with a drop of negative six point four nine percentage points in F1.

Jane: That tells us exactly where the model is focusing its attention when it makes these decisions, and that’s really helpful for interpretability.

Lu: The authors recommend this staged curriculum training because it gives you modular interpretability, spatial faithfulness, and few-shot responsiveness, which is how they achieved the results.

Meng: From a production standpoint, I'm leaning towards the modularity; if one part breaks, maybe we can isolate that concept module without crashing the whole classifier.

Title and authors: Tom: And to wrap up on this paper "GRAPE: Graph-Augmented Prototype Explanations for Interactive Medical Image Diagnosis," it really shows how you can move away from opaque outputs and toward a system where clinicians can actually verify or override the model's reasoning.

Jane: It’s about giving the doctor tools to check the math behind the AI, not just accepting whatever box it draws.

Lu: This work is significant because it moves beyond simple classification by explicitly modeling how anatomical concepts interact, which is a much richer way to understand medical images.

Meng: I still need to see if this entire four-stage training pipeline can actually be implemented on standard production hardware without needing massive retraining cycles every time a new condition pops up.

Tom: That's the key question for deployment, because they address the closed concept vocabulary problem by using Open-Vocabulary Prototype Anchoring to add new findings from just one labeled image.

Jane: So, we’ve seen how they use graphs to link concepts, how they build a safety net for physician feedback, and how they can grow the system without total retraining.

Lu: The overall implication is that by explicitly modeling concept co-occurrence and making prototypes adaptable through text alignment, we create a system that is both more accurate and far more interactive.

Meng: I'm still cautious about the real-world stability under unpredictable patient data, but the safety check mechanism does give me some confidence in its ability to catch immediate errors.

Tom: So, to finish up on this GRAPE paper, it’s a unified architecture that provides tangible improvements across classification accuracy and safety by addressing independent findings, unsafe feedback amplification, and the closed concept vocabulary problem.

Jane: It seems like the next step is seeing how this framework holds up when we move from controlled datasets like TBX11K to truly messy clinical data where every finding is unique.

Lu: That's definitely the next logical area for research, focusing on how robust those co-occurrence graphs are when they encounter novel or rare anatomical structures.

Meng: We'll have to see if the cost of training that graph structure is manageable before we can even think about scaling it up for a hospital environment.

Tom: Well, that’s our take on GRAPE today, showing how these architectural additions can provide concrete performance gains while also building in mechanisms for human verification.

Jane: We'll keep an eye on this space as researchers try to apply this framework to more complex diagnostic tasks.

The paper's summary: Tom: So, GRAPE is essentially taking those old prototype classifiers and giving them three major upgrades to fix their problems.

Jane: Right, so we talked about how they address independent concepts, unsafe feedback amplification, and needing a full retraining every time you want to add a new finding.

Tom: Exactly. The core idea is using this Graph Attention Task Head to model how anatomical concepts actually co-occur in the images.

Jane: And that's what’s really cool because it lets them create these prototype maps that are much better at localization, showing those gains over the baseline models.

Tom: Plus, they have this Concept-Mismatch Safety Check that actively warns a doctor if the model thinks one thing is dominant in an area but the doctor labeled it differently.

Jane: That safety mechanism sounds crucial for building trust; it stops bad feedback from silently poisoning the system's understanding of what's correct.

Tom: And they solve the open vocabulary issue with something called Open-Vocabulary Prototype Anchoring, which lets you add a new finding just by linking it to text descriptions.

Jane: So, what does this mean for a doctor using the tool? Does it give them more control over the AI’s reasoning than just seeing a final answer?

Tom: It gives them tools to verify or override the model's reasoning because every prediction is tied back to a specific visual concept and prototype.

Jane: That moves it from being just another black box output to something interactive where the human expert can actively check the logic.

Tom: And on speed, they made sure this whole graph thing adds only one millisecond of time when you run it at a batch size of one, which means it’s fast enough for some kind of real-time interaction.

Jane: It sounds like a lot of work to build all those stages—the concept supervision, the vector extraction, the prototype learning with that contrastive loss—but they're saying it’s worth it for modularity and responsiveness.

Tom: They stress that this staged training curriculum gives you modular interpretability, spatial faithfulness, and few-shot responsiveness.

Jane: So if you only listen to this paper, the big picture is that we can build systems where a doctor isn't just accepting an opaque output; they can actually verify or override the model’s reasoning using traceable visual concepts.

Tom: That traceability is what they call moving away from opaque outputs toward a system where clinicians check the math behind the AI.

Jane: It’s about giving those tools to check if the AI is looking at something it thinks it should, not just guessing based on pure statistical likelihood.

Tom: And they even showed that one specific edge in their graph—the connection to active tuberculosis—is actually the most important one for the model's performance.

Jane: That tells us exactly where the model is paying its attention, which helps with interpretation, and it shows how deeply these concepts are linked.

Tom: It’s a unified architecture that handles those three main flaws at once, making medical image diagnosis much more transparent and safer for potential deployment.

Jane: Now that we see how they fix the core structure, the next big question is whether this whole framework can actually handle the messy, unpredictable data you see in a real hospital setting.

The paper's improvements: Tom: So, we're looking at the specific improvements they are suggesting in this GRAPE architecture for medical imaging diagnosis.

Jane: Right, so we talked about the big three fixes—the graph head, the safety check, and open vocabulary anchoring—but what are the practical implications of those specific structural suggestions?

Tom: Well, one big suggestion is using a staged curriculum training pipeline instead of just one big training run.

Jane: That sounds like they're recommending a phased approach to learning the model’s structure before it gets fully optimized.

Tom: Exactly. They suggest starting with concept supervision, then extracting those concept vectors using the CAM weighted spatial averages of frozen features, and only then focusing on prototype learning with that loss L three.

Jane: That makes sense; it means they build a solid foundation for understanding concepts before trying to fine-tune the prototypes themselves.

Tom: And that structure is supposed to give you modular interpretability, spatial faithfulness, and few-shot responsiveness, which is what made them get those high F1 scores.

Jane: So they’re saying that by separating the learning stages like that, if one part needs tweaking—say, the concept extraction—you can adjust it without messing up the whole system.

Tom: That’s the engineering win; it makes debugging easier because you know which stage is responsible for which kind of error.

Jane: And on top of that, they emphasize that this structure allows for few-shot responsiveness, meaning adding a new finding doesn't require starting from scratch every single time.

Tom: That links back to the open vocabulary part; it shows the whole system is designed to grow incrementally rather than being rigidly fixed.

Jane: So, if you only listen to this section, remember that they are recommending a phased training approach because it helps you build a system that’s easier to trust and evolve over time.

Tom: And they also highlighted the role of the Graph Attention Task Head in boosting classification accuracy by thirteen point eight percentage points over their prototype baseline on that TBX11K dataset.

Jane: That number is concrete; it means the graph isn't just there for show, it’s actually providing a measurable performance benefit when you look at how anatomical concepts relate.

Tom: It’s a clean ablation showing exactly how much more accurate things get when you model co-occurrence this way.

Jane: So what this suggests for the future is that we might see more research focusing on these graph structures in other medical domains where understanding spatial relationships between findings matters a lot.

Conclusion: Tom: So to wrap up, GRAPE is this unified architecture that tackles independent findings, unsafe feedback amplification, and closed concept vocabulary in medical image diagnosis.

Jane: Exactly; it gives doctors a way to verify or override the model’s reasoning instead of just accepting whatever box it draws.

Tom: The implication is that we can build systems where clinicians check the math behind the AI, which is a big step toward safer deployment.

Jane: It’s about giving those tools to check if the AI is looking at something it thinks it should, not just guessing based on pure statistical likelihood.

Tom: This work shows how explicitly modeling concept co-occurrence and making prototypes adaptable through text alignment creates a system that is both more accurate and interactive.

Jane: The next big question we have to ask is how robust this framework holds up when you move from controlled datasets like TBX11K to the truly messy clinical data where every finding is unique.

Tom: That’s the challenge for future work; testing if those co-occurrence graphs are stable when they encounter novel or rare anatomical structures.

Jane: We also have to consider the practical side, Meng’s concern—does running this entire four-stage training pipeline cost too much in terms of resources or time?

Tom: Meng is right; we need to see if the cost of maintaining that graph structure is manageable before we even think about scaling it up for a hospital environment.

Jane: It seems like the next logical area for research is testing this framework on more complex diagnostic tasks where spatial relationships are just as important.

Tom: We've seen some solid wins here, but the real test will be in real-world environments, not just on benchmarks.

Jane: So that’s our wrap-up on GRAPE; it shows a powerful way to build systems that are both more accurate and much more interactive for human oversight.

Rasul Khanbayov, Erchin Serpedin, Hasan Kurban

Hamad Bin Khalifa University · Texas A&M University

cs.CV

Submitted: 2026-06-29

Updated: 2026-11-07

Code: https://github.com/KurbanIntelligenceLab/GRAPE

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 89/100

The gist: Prototype-based medical image classifiers present three clinical limitations: they treat findings as independent, silently amplify unsafe physician feedback…

Key concepts

Graph Attention Task Head
This module models how different anatomical concepts co-occur on an image using a data-driven graph. It improves classification accuracy by 13.8 percentage points over the baseline by capturing complex spatial relationships between findings.
Concept-Mismatch Safety Check
This check warns users when the model's dominant finding within a physician's drawn region contradicts its claimed label. It prevents silent amplification of unsafe feedback, catching 85% of erroneous annotations compared to other methods.
Open-Vocabulary Prototype Anchoring
This technique aligns visual prototypes with clinical text descriptions. It enables zero-shot concept addition, allowing new findings to be added from a single labeled image without needing to retrain or modify the existing model components.

Terminology

Summary

Prototype-based medical image classifiers present three clinical limitations: they treat findings as independent, silently amplify unsafe physician feedback, and require full retraining whenever a new finding is needed The authors present GRAPE (Graph-Augmented Prototype Explanations), a unified architecture that addresses all three challenges. p. 0

Where the paper says it

qa Abstract Prototype-based medical image classifiers present three clinical limitations: they treat findings as independent, silently amplify unsafe physician feedback, and require full retraining whenever a new finding is needed. We present GRAPE (Graph-Augmented Prototype Explanations), a unified architecture that addresses all three challenges.

Addressing Clinical Limitations

The limitations addressed by GRAPE include independent concepts, unsafe spatial interaction, and closed concept vocabulary. The proposed solution involves a Graph Attention Task Head to model anatomical concept co-occurrence a Concept-Mismatch Safety Check to warn when the model’s dominant finding inside a doctor-drawn region conflicts with the claimed label and Open-Vocabulary Prototype Anchoring to align visual prototypes to clinical text. p. 0

Where the paper says it

Second, a Concept-Mismatch Safety Check - the first such mechanism in prototype-based medical classifiers - warns when the model’s dominant finding inside a doctor-drawn region conflicts with the claimed label, catching 85% of erroneous annotations versus 51% for MC-Dropout with no extra inference cost. Third, Open-Vocabulary Prototype Anchoring aligns visual prototypes to clinical text, allowing a new finding to be added from a single labeled image without modifying any other component.

How the Training Signal is Built

The GRAPE training framework consists of a four-stage pipeline. Stage 1 involves concept supervision where the backbone and per-concept CAM heads are trained with BCE on Global Average Pooled activations. p. 3

Where the paper says it

Stage 1 — Concept Supervision. The backbone F and per-concept CAM heads are trained with BCE on Global Average Pooled activations (Appendix K, Eq.

Stage 2 extracts concept vectors by taking CAM-weighted spatial averages of frozen backbone features. p. 4
Where the paper says it

Stage 2 — Concept Vector Generation. CAM-weighted spatial averages of frozen backbone features produce per-concept vectors vk (Appendix K, Eq.

Stage 3 focuses on prototype learning, minimizing a loss L3 which includes the contrastive prototype learning loss Lcon and the VLM alignment loss Lalign. p. 4
Where the paper says it

Each projected concept vector v ′ k = P(vk) serves as a query; the multi-prototype contrastive loss assigns it to its ground-truth concept ˜k via soft assignment: q k m(v ′ ) = softmaxm(γ⟨pkm, v ′ ⟩), (10) simk(v ′ ) = X m q k m(v ′ )⟨pkm, v ′ ⟩, (11) Lcon = − log exp λ(simk˜(v ′ k˜ ) + δ) P k exp λ simk(v ′ k˜ ) , (12) with λ = 10, γ = 5, δ = 0. 1; the Stage 3 loss is L3 = Lcon + 0.

Stage 4 trains the Graph-Structured Classification head with a loss L4 = CE(GNN(s, G), y). p. 4
Where the paper says it

1; the Stage 3 loss is L3 = Lcon + 0. Stage 4 — Graph-Structured Classification.

Where the Gains Come From

The primary gains stem from the three core modules of GRAPE. The Graph Attention Task Head boosts macro-F1 by +13. p. 0

Where the paper says it

We present GRAPE (Graph-Augmented Prototype Explanations), a unified architecture that addresses all three challenges. First, a Graph Attention Task Head models anatomical concept co-occurrence, boosting macro-F1 by +13.

8 pp over the prototype baseline on TBX11K and the Concept-Mismatch Safety Check catches 85% of erroneous annotations versus 51% for MC-Dropout with no extra inference cost. p. 0
Where the paper says it

8 pp over the prototype baseline on TBX11K. Second, a Concept-Mismatch Safety Check - the first such mechanism in prototype-based medical classifiers - warns when the model’s dominant finding inside a doctor-drawn region conflicts with the claimed label, catching 85% of erroneous annotations versus 51% for MC-Dropout with no extra inference cost.

Furthermore, Open-Vocabulary Prototype Anchoring allows a new finding to be added from a single labeled image without modifying any other component. p. 0
Where the paper says it

Second, a Concept-Mismatch Safety Check - the first such mechanism in prototype-based medical classifiers - warns when the model’s dominant finding inside a doctor-drawn region conflicts with the claimed label, catching 85% of erroneous annotations versus 51% for MC-Dropout with no extra inference cost. Third, Open-Vocabulary Prototype Anchoring aligns visual prototypes to clinical text, allowing a new finding to be added from a single labeled image without modifying any other component.

Experimental Validation and Results

Experiments on TBX11K and NIH ChestX-ray14 cover classification, Pointing Game localization, inference speed, safety check validation, zero-shot concept addition, and Insertion/Deletion AUC faithfulness. GRAPE’s prototype maps achieve 2. p. 6

Where the paper says it

(3) Pointing Game and spatial faithfulness. GRAPE’s prototype maps achieve 2.

6× better lesion localization than end-to-end baselines on TBX11K and the safety check catches 85. p. 0
Where the paper says it

On NIH ChestX-ray14, one Effusion example recovers full-supervision localization accuracy; on TBX11K, prototype maps achieve 2. 6× better lesion localization than end-to-end baselines.

1% of misdrawn boxes with a FP rate of 16. p. 7
Where the paper says it

With the Uk-gated variant (θV = 2×10−4 ) as the default operating point, Module B catches 85. 1% of misdrawn boxes with an FP rate of 16.

4% using the Uk-gated variant. p. 7
Where the paper says it

Bootstrap 95% CIs use B = 1000 resamples. With the Uk-gated variant (θV = 2×10−4 ) as the default operating point, Module B catches 85.

Module C demonstrated that Effusion's 1-shot PG matches full supervision accuracy on NIH ChestX-ray14 (PG = 0.072, 100%). p. 7
Where the paper says it

We additionally evaluate on NIH ChestX-ray14 (K = 14): 3 of 4 target concepts have bbox coverage (Infiltration, Effusion, Atelectasis); Effusion’s 1-shot PG matches full supervision (PG = 0. 072, 100%), and Infiltration’s 20-shot reaches 53% of full supervision.

The Insertion/Deletion AUC evaluation confirms that prototype similarity maps capture genuinely predictive regions. p. 8
Where the paper says it

Insertion/Deletion AUC faithfulness evaluation (§4. 6) confirms that prototype similarity maps capture genuinely predictive spatial regions.

The GNN head adds only +1 ms over the linear head at batch size 1, satisfying real-time interaction requirements. p. 6
Where the paper says it

The edge ablation (Appendix G) shows self-loops dominate on TBX11K (K = 3); richer co-occurrence effects are expected on NIH (K = 14, 103 edges). The GNN head adds only +1 ms over the linear head at batch size 1, satisfying real-time interaction requirements (full latency benchmarks: Appendix F).

The GAT edge-importance ablation shows that the self-loop on concept 0 (ActiveTuberculosis) is the most informative edge, causing a drop of −6. p. 6
Where the paper says it

We retain the GAT head for its directly interpretable, learned edge-conditioned attention weights. Full results are in Appendix F; a per-edge importance ablation quantifying each edge’s contribution to final F1 is in Appendix G.

49 pp F1. The authors recommend the staged curriculum training for three reasons: modular interpretability, spatial faithfulness, and few-shot responsiveness. p. 7
Where the paper says it

Table 3: Summary comparison of GRAPE against the CSR baseline under the same interpretabilityconstrained training regime (frozen backbone, concept supervision, staged curriculum). All results are mean ± std over five seeds {0, 1, 2, 3, 42}; significance via paired t-test on per-seed Macro F1.

The work concludes that GRAPE enables clinicians to verify or override model reasoning rather than accepting opaque outputs. p. 8
Where the paper says it

Broader Impact Interpretable AI systems for medical imaging have the potential to meaningfully improve clinical practice. By making every prediction traceable to a specific visual concept and prototype, GRAPE enables clinicians to verify or override model reasoning rather than accepting opaque outputs.

The system is a research prototype and is not deployed in any clinical setting. p. 8
Where the paper says it

All experiments are conducted retrospectively on de-identified images. The system described in this paper is a research prototype and is not deployed in any clinical setting.

The authors acknowledge the use of an NVIDIA A100 80GB GPU for these experiments. p. 9
Where the paper says it

Acknowledgments and Disclosure of Funding The authors thank the TBX11K and NIH ChestX-ray14 dataset providers for making their data publicly available. This work used an NVIDIA A100 80GB GPU.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Graph Attention Task Head improves classification accuracy by +13.8 pp over prototype baseline on TBX11K by modeling anatomical concept co-occurrence via a data-driven co-occurrence graph.

  2. Concept Mismatch Safety Check prevents silent error amplification when physicians provide corrective feedback by warning when the model’s dominant finding inside a doctor-drawn region conflicts with the claimed label, catching 85% of erroneous annotations compared to 51% for MC-Dropout.

  3. Open Vocabulary Prototype Anchoring enables zero-shot concept addition by aligning visual prototypes to clinical text descriptions via a frozen BioViL-T encoder, allowing a new finding to be added from a single labeled image without modifying any other component.

  4. The system can perform interactive medical image diagnosis where clinicians can verify or override model reasoning because every prediction is traceable to a specific visual concept and prototype.

  5. The system reduces the risk of harmful feedback degrading the prototype atlas by using the concept-mismatch check, which is critical for clinical trust in deployment scenarios.

Abstract

Prototype-based medical image classifiers present three clinical limitations: they treat findings as independent, silently amplify unsafe physician feedback, and require full retraining whenever a new finding is needed. We present GRAPE (Graph-Augmented Prototype Explanations), a unified architecture that targets all three challenges. First, a Graph Attention Task Head boosts macro-F1 by +13.8,pp over the prototype baseline on TBX11K but not on NIH ChestX-ray14, with no graph-specific benefit established. Second, a Concept-Mismatch Safety Check, to our knowledge the first such mechanism in prototype-based medical classifiers, warns when the model's dominant finding inside a doctor-drawn region conflicts with the claimed label, catching 85% of simulated erroneous annotations with no extra inference cost at a favorable operating point, though sensitivity falls substantially at clinically realistic false-positive budgets. Third, Open-Vocabulary Prototype Anchoring aligns visual prototypes to clinical text, allowing a new finding to be added from a single labeled image without modifying any other component, evaluated for localization only. On NIH ChestX-ray14, one Effusion example matched that concept's low fully-supervised Pointing Game ceiling in one run but not in a second. On TBX11K, prototype maps achieve 2.1x better lesion localization than end-to-end concept-based baselines. The graph head adds only +1 ms latency at interactive batch size. The project page is available at https://kurbanintelligencelab.github.io/GRAPE.

Sources

Related papers