Research map / Graph learning / Knowledge and temporal graphs

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

NeurIPS 2026 · Graph learning › Knowledge and temporal graphs · approach: Graph reasoning for visual recognition · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GRAPE, a prototype-based medical image classifier that adds a GAT task head over label co-occurrence, a prototype-variance/concept-mismatch safety check for doctor-drawn boxes, and BioViL-T text anchoring for zero-shot concept addition, evaluated on TBX11K and NIH ChestX-ray14.

Problem

Prototype-based medical image classifiers treat findings as independent, amplify unsafe physician spatial feedback, and need full retraining to add a new finding.

Approach

Combines a graph attention head over a training-label co-occurrence graph on prototype similarity scores, a deterministic concept-mismatch check on doctor-drawn boxes using prototype mean maps, and alignment of prototypes to frozen BioViL-T text embeddings so new concepts can be added from text or one image.

Key results

• GRAPE improves TBX11K macro F1 from 0.728 (CSR) to 0.866 over 5 seeds (+13.8 pp, p=0.0013).
• Concept-mismatch safety check with U_k gating catches 85.1% of simulated erroneous boxes at 16.4% FP rate, versus 51.0% TP for MC-Dropout.
• On NIH, GRAPE macro F1 (0.6031) is statistically indistinguishable from CSR (0.6048; p=0.894) under the multi-seed patient-level split.

Surprising findings

• The multi-seed Pointing Game advantage over CSR does not replicate (p=0.64), despite a single-seed +7.73 pp gain.
• End-to-end training reaches 0.990 F1 with a plain ResNet-50 linear head at 0.993, so F1 gains mostly reflect unconstrained backbone fine-tuning rather than concept reasoning.

Limitations

• Pointing Game scores are low (7-14%) due to 7x7 similarity-map resolution, and multi-seed PG differences are not significant.
• NIH multi-seed results show no F1 gain over CSR; the NIH safety check has high FP rate (0.731) due to dense concept overlap.
• Zero-shot Module C validated only on 3 concepts with bbox coverage on NIH; no clinician user study.

Experimental setup

Datasets: TBX11K NIH ChestX-ray14

Benchmarks: TBX11K NIH ChestX-ray14

Metrics: macro F1 pointing game hit rate true-positive rate false-positive rate insertion AUC deletion AUC latency

Baselines: CSR CBM PIPNet ML-GCN ADD-GCN MC-Dropout ResNet-50

Builds on: CSR ProtoPNet BioViL-T GAT ML-GCN

Closest papers by meaning

• Interactive Medical Image Analysis with Concept-based Similarity Reasoning (CVPR 2025 · similarity 0.88)
• CXReasonBench: A Benchmark for Evaluating Structured Diagnostic Reasoning in Chest X-rays (NeurIPS 2025 · similarity 0.86)
• M3CoTBench: Benchmark Chain-of-Thought of MLLMs in Medical Image Understanding (ICLR 2026 · similarity 0.85)
• CRAFT: Causal Responsibility and Failure Tracing in Medical Vision Language Models (NeurIPS 2026 · similarity 0.85)
• Rethinking Visual Attribution for Chest X-ray Reasoning in Large Vision Language Models (NeurIPS 2026 · similarity 0.85)
• OpenMedReason: Scientific Reasoning Supervision for Medical Vision–Language Models (NeurIPS 2026 · similarity 0.85)