Learning to Trim: End-to-End Causal Graph Pruning with Dynamic Anatomical Feature Banks for Medical VQA

Dimensions overall score 7.0

• The proposed Learnable Causal Trimming (LCT) framework integrates causal pruning into end-to-end optimization for Medical VQA.method 95%
  Evidencepartial

  To address this issue, we propose a Learnable Causal Trimming (LCT) framework that integrates causal pruning into end-to-end optimization.

  Implicationpartial

  This is a core statement of the proposed method, explicitly mentioned in the abstract and introduction.

  Verificationpartial

  partial

• A Dynamic Anatomical Feature Bank (DAFB) is introduced, updated via a momentum mechanism, to capture global prototypes of frequent anatomical and linguistic patterns.method 95%
  Evidencepartial

  We introduce a Dynamic Anatomical Feature Bank (DAFB), updated via a momentum mechanism, to capture global prototypes of frequent anatomical and linguistic patterns, serving as an approximation of dataset-level regularities.

  Implicationpartial

  This describes a key component of the proposed method, clearly stated in the abstract.

  Verificationpartial

  partial

• The LCT framework consistently improves robustness and generalization over existing debiasing strategies on VQA-RAD, SLAKE, SLAKE-CP, and PathVQA datasets.result 95%
  Evidencepartial

  Experiments on VQA-RAD, SLAKE, SLAKE-CP and PathVQA demonstrate that LCT consistently improves robustness and generalization over existing debiasing strategies.

  Implicationpartial

  This is a major result claim, directly stated in the abstract and supported by multiple tables showing performance improvements.

  Verificationpartial

  partial

• LCT achieves state-of-the-art performance and improves Out-of-Distribution (OOD) generalization on four datasets.result 95%
  Evidencepartial

  Experiments on four datasets demonstrate that LCT achieves state-of-the-art performance and improves OOD generalization.

  Implicationpartial

  This is a strong claim about the performance and generalization capabilities of the proposed method, stated in the introduction and supported by results tables.

  Verificationpartial

  partial

• LCT achieves the highest Overall accuracy of 65.6% on the PathVQA dataset, outperforming MUMC (65.1%).result 90%
  Evidencepartial

  To further assess cross-modality generalization, we evaluate on PathVQA (Table 2). LCT achieves the highest Overall accuracy of 65.6%, outperforming MUMC (65.1%).

  Implicationpartial

  Specific numerical result comparing LCT to a baseline on a particular dataset.

  Verificationpartial

  partial

• LCT establishes a new SOTA on Open-ended questions for PathVQA with an accuracy of 41.1%.result 90%
  Evidencepartial

  Notably, LCT establishes a new SOTA on Open-ended questions (41.1%), while achieving comparable Closed accuracy (89.9%).

  Implicationpartial

  Specific numerical result highlighting a SOTA achievement on a subset of questions for a particular dataset.

  Verificationpartial

  partial

• LCT achieves the best performance across all metrics on SLAKE-CP, reaching 51.0% Overall accuracy.result 90%
  Evidencepartial

  Table 3 presents results on SLAKE-CP, a challenging OOD benchmark for evaluating diagnostic robustness. LCT achieves the best performance across all metrics, reaching 51.0

  Implicationpartial

  Specific numerical result indicating best performance on an OOD benchmark.

  Verificationpartial

  partial

• With LCT, Grad-CAM visualizations show more anatomically focused attention maps, concentrating on relevant structures.technical 85%
  Evidencepartial

  With LCT, the attention maps become more anatomically focused. In the “w/ LCT” column, activations concentrate on the relevant structures, such as the small nodule in the left lung (Example a) and the enlarged cardiac silhouette (Example b), indicating improved visual grounding.

  Implicationpartial

  This describes a qualitative improvement in model interpretability and grounding, supported by visual examples and textual explanation.

  Verificationpartial

  partial

• The proposed Learnable Causal Trimming (LCT) framework integrates causal pruning into end-to-end optimization for Medical VQA.method 95%
  Evidencepartial

  To address this issue, we propose a Learnable Causal Trimming (LCT) framework that integrates causal pruning into end-to-end optimization.

  Implicationpartial

  This is a core statement of the paper's proposed method, explicitly mentioned in the abstract and introduction.

  Verificationpartial

  partial

• The Dynamic Anatomical Feature Bank (DAFB) is updated via a momentum mechanism to capture global prototypes of frequent anatomical and linguistic patterns.method 95%
  Evidencepartial

  We introduce a Dynamic Anatomical Feature Bank (DAFB), updated via a momentum mechanism, to capture global prototypes of frequent anatomical and linguistic patterns, serving as an approximation of dataset-level regularities.

  Implicationpartial

  This describes a key component of the proposed method and its function, as stated in the abstract.

  Verificationpartial

  partial

• The LCT framework consistently improves robustness and generalization over existing debiasing strategies on VQA-RAD, SLAKE, SLAKE-CP, and PathVQA datasets.result 95%
  Evidencepartial

  Experiments on VQA-RAD, SLAKE, SLAKE-CP and PathVQA demonstrate that LCT consistently improves robustness and generalization over existing debiasing strategies.

  Implicationpartial

  This is a primary result claimed in the abstract and supported by multiple tables showing performance improvements.

  Verificationpartial

  partial

• LCT achieves state-of-the-art performance and improves Out-of-Distribution (OOD) generalization on four datasets.result 95%
  Evidencepartial

  Experiments on four datasets demonstrate that LCT achieves state-of-the-art performance and improves OOD generalization.

  Implicationpartial

  This is a strong claim about the model's performance, explicitly stated in the introduction and supported by results tables.

  Verificationpartial

  partial

Competitive landscape

A framework for medical visual question answering that learns to suppress spurious correlations and focus on causal diagnostic evidence, improving generalization.