Rethinking Visual Attribution for Chest X-ray Reasoning in Large Vision Language Models

Technical Summary: Rethinking Visual Attribution for Chest X-ray Reasoning in Large Vision Language Models

Problem Statement
Large Vision Language Models (LVLMs) are increasingly deployed in high-stakes medical scenarios, such as radiology report generation and diagnostic assistance. However, a critical barrier to clinical trustworthiness is the inability to faithfully attribute model outputs to specific visual evidence in the input. While various visual attribution methods (gradient-based, attention-based, perturbation-based, and prompting-based) exist, their reliability in medical settings remains unverified. The core challenge is the absence of objective ground truth for internal model reasoning; human annotations often reflect clinical relevance rather than the specific visual cues a black-box model actually used to make a decision. Consequently, existing attribution methods may highlight regions that are clinically plausible but causally irrelevant to the model's prediction, or conversely, miss the true evidence entirely.

Methodology
The authors propose a two-pronged approach: a rigorous causal evaluation framework and a new attribution method, MedFocus.

1. Causal Evaluation Framework (MedGround-Bench):
   To objectively evaluate attribution faithfulness, the authors constructed MedGround-Bench, a benchmark derived from three chest X-ray (CXR) datasets (ImaGenome, VinDR-CXR, PadChest-GR) containing expert-annotated spatial regions. The framework employs a three-step causal filtering process to retain only samples where the expert-annotated region is verified to be causally responsible for the model's prediction:

   • Correctness Filtering: Retain only samples where the LVLM answers the binary VQA question ("Is there evidence of [attribute]?") correctly.
   • Foreground Counterfactual Editing: Using the RadEdit model, the annotated region is inpainted to remove the target attribute. Samples are retained only if the model's answer flips (e.g., from "Yes" to "No"), confirming the region's causal necessity.
   • Background Counterfactual Editing: The background (outside the annotated region) is edited. Samples are retained only if the model's answer remains unchanged, ensuring the decision change in the foreground step was not an artifact of general image sensitivity.
     This process yields a curated set of 3,940 samples across six LVLMs (including Qwen2.5-VL, Gemma3, and MedGemma) and two output modes (direct answer and step-by-step reasoning).

2. MedFocus Attribution Method:
   To address the failure of existing methods, the authors propose MedFocus, a concept-based causal attribution method. Unlike pixel-level saliency approaches, MedFocus operates on clinically meaningful anatomical concepts (e.g., left lung, cardiac silhouette).

   • Concept Segmentation: It localizes anatomical regions using Unbalanced Optimal Transport (UOT) to map a reference normal CXR with known annotations to the target image, accommodating pathological mass variations. These initial masks are refined using MedSAM for clean boundaries.
   • Causal Intervention: For each concept, MedFocus generates a counterfactual by zero-masking the concept's bounding box (rather than the pixel mask) and measures the cumulative drop in token-level log-probabilities of the original output sequence.
   • Output: It produces spatial attributions (bounding boxes), concept-level explanations (anatomical names), and token-level attributions (identifying which reasoning tokens are most affected by the intervention).

Key Results

• Failure of Existing Methods: Evaluation on MedGround-Bench reveals that 11 existing attribution methods (including GradCAM, Integrated Gradients, Attention Rollout, and prompting-based approaches) consistently fail to faithfully identify the visual evidence driving LVLM predictions. They often produce diffuse maps with low precision or focused maps that miss the true evidence, regardless of the model architecture or output mode.
• MedFocus Performance: MedFocus substantially outperforms all baselines across all datasets and models. In the direct answer mode, it achieves an IoU of 54.24% on ImaGenome compared to the next best (Gradient-weighted Attention at 39.24%). It maintains strong performance in the reasoning mode, where other methods show significant degradation.
• Model Behavior Insights: The study finds that medically trained models (MedGemma) exhibit higher attribution faithfulness and lower failure rates than generalist models. Furthermore, models generating step-by-step reasoning tend to engage more consistently with anatomically meaningful evidence than those providing direct answers.
• Ablation Studies: The two-stage segmentation pipeline (UOT + MedSAM) and the use of bounding-box zero-masking for intervention were identified as critical design choices that outperform end-to-end segmentation or pixel-level masking.

Significance and Claims
The paper claims that current attribution methods are insufficient for ensuring the trustworthiness of medical LVLMs, as they frequently fail to align with the actual visual evidence used by the model. By introducing a causally validated benchmark and a concept-based attribution method, the authors provide:

1. Evidence of a Gap: A demonstration that existing techniques do not reliably ground medical LVLM predictions in visual evidence, even when the model answers correctly.
2. A Solution: MedFocus, which bridges low-level visual evidence and high-level clinical understanding by grounding explanations in interpretable anatomical concepts, offering spatial, concept-level, and token-level attributions that are more faithful and clinically interpretable.
3. A Resource: The release of MedGround-Bench to support rigorous, causally grounded evaluation of attribution methods in future medical AI research.

The authors position this work as a step toward more trustworthy medical LVLMs, emphasizing that reliable attribution is essential for clinician trust, error detection, and patient safety, while acknowledging that their current framework is specific to CXR due to data availability constraints.