Attention Without Grounding: Causal Evaluation of Visual Explanations in Medical VLMs

• URL: http://arxiv.org/abs/2607.18577v1
• Date: Mon, 20 Jul 2026 23:19:28 GMT
• Title: Attention Without Grounding: Causal Evaluation of Visual Explanations in Medical VLMs
• Abstract summary: Attention and saliency heatmaps are widely used to explain medical Vision-Language Model (VLM) outputs on chest X-rays.<n>We audit faithfulness via overlap with radiologist bounding boxes on PadChest.<n>We study three MedGemma-4B variants, cross-family probes on LLaVA-RAD and Qwen3-VL-8B-Instruct, and the specialist CheXagent-2-3b.
• Score: 2.064612766965483
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: Attention and saliency heatmaps are widely used to explain medical Vision-Language Model (VLM) outputs on chest X-rays, yet whether they truly highlight the image evidence driving predictions has not been causally tested. We audit faithfulness via overlap with radiologist bounding boxes on PadChest (n=637), attribution mass within radiologist masks on CheXlocalize (n=643), and 16x16 patch-occlusion maps that record which regions, when hidden, change the answer. We study three MedGemma-4B variants, cross-family probes on LLaVA-RAD and Qwen3-VL-8B-Instruct, and the specialist CheXagent-2-3b, with two CXR-trained classifiers (DenseNet121, ResNet50) as positive controls. A heatmap is faithful only if the model uses the image and attention concentrates on regions whose occlusion alters the prediction. No evaluated VLM meets both criteria. MedGemma and Qwen3-VL use the image, but attention anti-correlates with patch-occlusion importance (rho < 0 with 95% bootstrap CIs below zero). LLaVA-RAD's attention correlates positively, but the model is almost text-only (99.1% text-only agreement, near-zero causal mass), so correlation ties two near-zero signals. Attention also misses annotated anatomy: overlap with true regions never beats shifted or random controls, and no method places more than 22% of its mass inside radiologist masks. The two CXR classifiers pass all metrics, indicating the failure is specific to VLM heatmaps, not the evaluation. These heatmaps are visually reassuring but not faithful; clinical explanations require controlled localization metrics and causal perturbation, not visual inspection alone.

Related papers

• Pathology Transport: Optimal-Transport Explanations for Clinical Data, and When Their Heatmaps (Fail to) Localize Disease [0.9518285745183249]
  We build an optimal-transport rectified flow trained between two clinical distributions.<n>A single flow yields per-patient counterfactuals, an unsupervised malignancy score and a label-free attribution.<n>On chest X-rays, we show the transport heatmap is a population-level signal, not a localiser.
  arXiv  Detail & Related papers  (2026-08-18T04:59:46Z)
• A Controlled Audit of Pretraining Contamination in Public Medical Vision-Language Benchmarks [46.43161388634735]
  Medical vision-language models (VLMs) are evaluated on public benchmarks whose images and question-answer pairs have been freely downloadable for years.<n>We audit open VLMs on SLAKE-En, PathVQA, VQA-RAD, and an auxiliary public OmniMedVQA mirror using four detector families.
  arXiv  Detail & Related papers  (2026-06-08T18:40:10Z)
• CXR-ContraBench: Benchmarking Negated-Option Attraction in Medical VLMs [20.96410413299322]
  CXR-ContraBench is a diagnostic benchmark spanning internal ReXVQA slices and external OpenI and CheXpert protocols.<n>We study this failure as negated-option attraction, where a model is drawn to a negated answer option even when it conflicts with both the visual evidence and the question.
  arXiv  Detail & Related papers  (2026-05-07T07:46:17Z)
• Wasserstein-Aligned Localisation for VLM-Based Distributional OOD Detection in Medical Imaging [8.915586506306875]
  We introduce WALDO, a training-free framework grounded in optimal transport theory that enables comparative reasoning.<n>We show that WALDO with Qwen2.5-VL-72B achieves $43.5_pm1.6%$ mAP@30 (95% CI: [40.4, 46.7]), representing a 19% relative improvement over zero-shot baselines.
  arXiv  Detail & Related papers  (2026-05-06T17:32:34Z)
• Evaluating Reasoning Faithfulness in Medical Vision-Language Models using Multimodal Perturbations [19.488236277427358]
  Vision-language models (VLMs) often produce chain-of-thought (CoT) explanations that sound plausible yet fail to reflect the underlying decision process.<n>We present a clinically grounded framework for chest X-ray visual question answering (VQA) that probes CoT faithfulness via controlled text and image modifications.
  arXiv  Detail & Related papers  (2025-10-13T09:28:22Z)
• DualAttNet: Synergistic Fusion of Image-level and Fine-Grained Disease Attention for Multi-Label Lesion Detection in Chest X-rays [1.3367903535457364]
  We propose a dual attention supervised module for multi-label lesion detection in chest radiographs, named DualAttNet. It efficiently fuses global and local lesion classification information based on an image-level attention block and a fine-grained disease attention algorithm.
  arXiv  Detail & Related papers  (2023-06-23T23:19:27Z)
• Attention-based Saliency Maps Improve Interpretability of Pneumothorax Classification [52.77024349608834]
  To investigate chest radiograph (CXR) classification performance of vision transformers (ViT) and interpretability of attention-based saliency. ViTs were fine-tuned for lung disease classification using four public data sets: CheXpert, Chest X-Ray 14, MIMIC CXR, and VinBigData. ViTs had comparable CXR classification AUCs compared with state-of-the-art CNNs.
  arXiv  Detail & Related papers  (2023-03-03T12:05:41Z)
• Radiomics-Guided Global-Local Transformer for Weakly Supervised Pathology Localization in Chest X-Rays [65.88435151891369]
  Radiomics-Guided Transformer (RGT) fuses textitglobal image information with textitlocal knowledge-guided radiomics information. RGT consists of an image Transformer branch, a radiomics Transformer branch, and fusion layers that aggregate image and radiomic information.
  arXiv  Detail & Related papers  (2022-07-10T06:32:56Z)
• Explainable multiple abnormality classification of chest CT volumes with AxialNet and HiResCAM [89.2175350956813]
  We introduce the challenging new task of explainable multiple abnormality classification in volumetric medical images. We propose a multiple instance learning convolutional neural network, AxialNet, that allows identification of top slices for each abnormality. We then aim to improve the model's learning through a novel mask loss that leverages HiResCAM and 3D allowed regions.
  arXiv  Detail & Related papers  (2021-11-24T01:14:33Z)
• Explaining COVID-19 and Thoracic Pathology Model Predictions by Identifying Informative Input Features [47.45835732009979]
  Neural networks have demonstrated remarkable performance in classification and regression tasks on chest X-rays. Features attribution methods identify the importance of input features for the output prediction. We evaluate our methods using both human-centric (ground-truth-based) interpretability metrics, and human-independent feature importance metrics on NIH Chest X-ray8 and BrixIA datasets.
  arXiv  Detail & Related papers  (2021-04-01T11:42:39Z)
• Contralaterally Enhanced Networks for Thoracic Disease Detection [120.60868136876599]
  There exist many similar structures in the left and right parts of the chest, such as ribs, lung fields and bronchial tubes. This kind of similarities can be used to identify diseases in chest X-rays, according to the experience of broad-certificated radiologists. We propose a deep end-to-end module to exploit the contralateral context information for enhancing feature representations of disease proposals.
  arXiv  Detail & Related papers  (2020-10-09T10:15:26Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.