Source-linked AI summary
Beyond Explanation: Debugging Medical Imaging Models via Concept Intervention
Samrajya Thapa, Daniel J. Quest, Timothy L. Kline, Carrie L. Langstraat, Emanuel C. Trabuco, Wei Le
TL;DR
Medical imaging models can be difficult to interpret and debug because their predictions lack explicit links to clinical concepts. This paper aligns single-modality encoders with BioMedCLIP to enable concept-level interventions and targeted refinement. On the evaluated datasets, it reports maintained baseline performance and 2–3% downstream gains.
Problem
Black-box medical imaging encoders lack explicit links between predictions and clinically meaningful abstractions, complicating trust, validation, and systematic error analysis.
Source: §1 · pp. 1–2Method
The framework linearly aligns a task-specific image encoder to BioMedCLIP, builds a concept bottleneck, and applies concept interventions to diagnose and refine predictions.
Source: §1 · p. 2, §3 · p. 3, §3.1 · p. 4Results
The framework maintains baseline performance and boosts downstream capabilities by 2–3% through targeted intervention.
Source: §1 · p. 2Takeaways & Limitations
Intervenable concept bottlenecks offer a plug-in route to concept-based debugging and controlled refinement of off-the-shelf medical imaging models.
Source: §5 · pp. 8–9Radiology narratives describe global multi-view studies and lack granular association with individual images or slices, constraining text-to-image concept alignment.
Source: §5 · p. 9
Abstract
from arXiv · showhide
Medical imaging models often operate as black boxes, limiting interpretability and systematic debugging. We introduce an easy-to-use, plug-and-play framework for concept-based interpretation and model refinement. By aligning a single-modality encoder to BioMedCLIP, we construct a Concept Bottleneck Model (CBM) that enables concept-level interventions. These interventions allow us to isolate causal versus spuriously correlated concepts, validate insights with domain experts, and generate counterfactual samples for targeted fine-tuning. We evaluate our framework on a Mayo Clinic ultrasound dataset and the CheXpert 5x200 chest X-ray dataset. Results demonstrate that concept intervention enables reliable model diagnosis while maintaining, and occasionally improving predictive performance via guided fine-tuning. Our findings highlight the practical value of this framework for controlled, interpretable refinement of clinical deep learning models.
1 Introduction
Medical imaging models often lack links between predictions and clinically meaningful concepts, complicating trust, validation, and error analysis. The paper proposes an aligned concept bottleneck framework for debugging and refining these models, reporting maintained baseline performance and 2–3% downstream gains.
1 Introduction
- Black-box image encoders lack explicit links between predictions and clinically meaningful abstractions, complicating trust, validation, and systematic error analysis.The introduction identifies interpretability as important for trust between clinicians and AI models.Source: §1 · p. 1
- Post-hoc visual explanations lack semantic grounding, while standard concept bottleneck models often require dense concept supervision and end-to-end training.These requirements limit standard CBMs for pretrained, frozen, or deployed models.Source: §1 · pp. 1–2
- The proposed framework linearly aligns a task-specific encoder with BioMedCLIP and uses concept interventions to identify missing, weak, or spurious concepts.Its contributions include expert validation and intervention-guided counterfactual refinement of the classification head.Source: §1 · p. 2
- The framework maintains baseline performance and boosts downstream capabilities by 2–3% through targeted interventions on Mayo ultrasound and CheXpert 5×200.The interventions use clinical concepts to help models focus on diagnostic features.Source: §1 · p. 2
2 Related Work
The related work contrasts concept-based explanations with pixel-level attribution, then identifies gaps in post-hoc CBMs and intervention-guided refinement. The paper positions its approach as a lightweight alignment and debugging framework for single-modality medical models.
Feature Attribution vs. Concept Interpretability
- Pixel-level heatmaps can lack semantic grounding and vary with architectures or small input perturbations, whereas concept methods route predictions through explicit semantic representations.Concept methods use human-understandable representations rather than post-hoc interpretation of pixel-level features.Source: §2 · p. 3
Vision-Language Alignment and Post-Hoc CBMs
- Standard CBMs require dense annotations and end-to-end retraining, while vision–language methods enable post-hoc mapping into textual concept spaces.The paper builds on this paradigm with a lightweight affine mapping into a clinically grounded concept space.Source: §2 · p. 3
Concept Intervention and Model Refinement
- Prior concept interventions probe counterfactual states, but expert-guided use to debug and fine-tune frozen single-modality medical models remains underexplored.The proposed work turns test-time interventions into latent data augmentation and classification-head refinement.Source: §2 · p. 3
3 Methodology
The methodology aligns a single-modality image encoder with BioMedCLIP, constructs a concept bottleneck, and uses counterfactual interventions for diagnosis and targeted refinement. It also describes concept selection and expert validation across ultrasound and chest X-ray data.
3 Methodology
- The pipeline combines post-hoc feature mapping, concept bottleneck construction, and counterfactual intervention for diagnosis and downstream refinement.The overview describes intervention-guided concept perturbations as the basis for targeted augmentation and classification-head fine-tuning.Source: §3 · p. 3, Figure 1 · p. 4
3.1 Post-Hoc Alignment and CBM Construction
- Image features are extracted before the final classification layer and mapped by an affine transformation into BioMedCLIP’s embedding space.The features come after global average pooling; the alignment is presented as a proof of concept given global radiology notes and image-frame mismatch.Source: §3.1 · p. 4
3.2 Counterfactual Concept Intervention and Diagnosis
- Concept interventions overwrite selected predicted activations with alternative binary states, and gradient-guided ranking prioritizes concepts for finding a prediction-flipping set.The greedy search ranks concepts by absolute gradient magnitude with respect to the target-class logit.Source: §3.2 · p. 5
3.3 Intervention-Guided Refinement
- Successful interventions on erroneous predictions create a counterfactual augmentation dataset, while the image encoder and alignment matrix remain frozen during classification-head fine-tuning.The latent bottleneck avoids pixel-level image synthesis; fine-tuning is restricted to the task classification head.Source: §3.3 · pp. 5–6
3.4 Concept Selection and Expert Validation
- The study curates 34 Mayo concepts from clinical metadata and 10 CheXpert concepts from paired reports, then evaluates concept predictions and interventions with expert or report-based checks.Mayo concepts are curated with domain experts; CheXpert report matching uses regex with strict negation detection.Source: §3.4 · p. 6
4 Results
Across ultrasound and CheXpert evaluations, intervention-guided CBMs preserve or improve predictive performance while providing concept-level interpretability. Case studies also show counterfactual interventions correcting errors associated with spurious concepts.
4 Results
- Evaluation covers a 525-sample Mayo Clinic ovarian ultrasound dataset, an additional 644-sample out-of-distribution validation set, and CheXpert 5×200 with 1,000 chest X-rays across five findings.The CheXpert evaluation uses a pretrained DenseNet backbone from TorchXrayVision.Source: §4 · p. 6
- On CheXpert, the aligned zero-shot model reaches 33.3% accuracy versus 33.9% for the CLIP baseline, with AUROC of 64.1% versus 62.3%.The authors characterize the projection as retaining downstream representation quality within a 1% absolute performance margin.Source: §4 · p. 7
- On ultrasound, the CBM preserves baseline task accuracy in-distribution, while counterfactual refinement improves accuracy and F1-score on the out-of-distribution test split.Post-intervention fine-tuning yields only a marginal AUROC edge on the in-distribution split.Source: §4 · p. 7
- On CheXpert, post-intervention fine-tuning improves macro-averaged recall, F1-score, and AUROC after the pre-intervention CBM balances precision and recall relative to baseline.The pre-intervention CBM addresses baseline sensitivity flaws, including on Consolidation; Table 2 compares per-label performance before and after intervention-based fine-tuning.Source: §4 · pp. 7–8, Table 2 · p. 7
- The intervention-guided CBM achieves the highest overall accuracy and Macro F1 among evaluated CheXpert frameworks while providing intrinsic interpretability.Table 3 compares overall performance across models on CheXpert 5×200.Source: Table 3 · p. 7, §4 · p. 8
- Case studies show concept interventions correcting false predictions on ovarian masses and resolving multi-label ambiguities in chest X-rays.The examples span Mayo Clinic ultrasound and CheXpert 5×200.Source: §4 · p. 8, Figure 2 · p. 8
5 Conclusion
The paper presents a plug-in concept bottleneck framework for interpreting and refining off-the-shelf medical imaging models. It identifies a limitation in linking global clinical narratives to specific image frames and describes potential auditing uses in clinical workflows.
5 Conclusion
- Linear mapping to BioMedCLIP creates an intervenable concept bottleneck for off-the-shelf medical models without architectural modifications.Gradient-guided concept interventions support counterfactual debugging and targeted classification-head fine-tuning.Source: §5 · pp. 8–9
- The authors identify a spatial and structural disconnect between global clinical narratives and specific image frames as a key limitation.Radiology summaries cover multi-view studies globally and lack granular instance-level links to images or slices.Source: §5 · p. 9
- The proposed application is auditing legacy clinical workflows through realtime, human-in-the-loop debugging and diagnostic-head refinement at the point of care.The passage describes these uses as possible deployment prospects, without end-to-end retraining.Source: §5 · p. 9