Research map / Label-efficient and robust learning / Recognition and long-tail
Towards Multi-dimensional Explanation Alignment for Medical Classification
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 Med-MICN, an end-to-end concept-bottleneck framework for medical image classification that auto-labels concepts with GPT-4V and BioViL, and combines concept embeddings, neural-symbolic logic rules, and saliency maps for multi-dimensional explanations.
Problem
Medical image classifiers lack interpretability; existing concept-based models depend on expensive concept annotations, ignore intrinsic image features (hurting accuracy), and their concept importances are misaligned with saliency explanations.
Approach
Generate concept sets with GPT-4V, pseudo-label concepts per image using BioViL image-text similarity heatmaps, then jointly train a concept encoder, a neural-symbolic fuzzy-logic rule layer, and a classifier that fuses concept embeddings with backbone features, yielding concept scores, logic rules, and saliency maps as aligned explanations.
Key results
- On COVID-CT with ResNet50 backbone, Med-MICN reaches 84.75% accuracy versus 81.36% for the ResNet50 baseline; with DenseNet169 it reaches 86.44%.
- On DDI with VGG19, accuracy rises 6.01 points over the baseline (to 82.58%), and on Chest X-Ray VGG19 reaches 88.30% accuracy.
- Ablation on DDI shows L_c alone raises AUC by 3.79 points and both losses together raise AUC by 8.71 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Med-MICN instead of the strongest backbone baseline | DenseNet169 · COVID-CT | accuracy: 85.59 -> 86.44 | Helped |
| use Med-MICN instead of the strongest backbone baseline | VGG19 · DDI | accuracy: 78.03 -> 82.58 | Helped |
| use Med-MICN instead of the strongest backbone baseline | VGG19 · Chest X-Ray | accuracy: 81.41 -> 88.30 | Helped |
| use Med-MICN instead of the strongest backbone baseline | ResNet50 · Fitzpatrick17k | accuracy: 80.79 -> 82.76 | Helped |
| add concept loss (ablation) | Med-MICN with neural-symbolic loss only · COVID-CT | accuracy: 81.36 -> 84.75 | Helped |
| add neural-symbolic loss (ablation) | Med-MICN with concept loss only · COVID-CT | accuracy: 83.05 -> 84.75 | Helped |
| add concept loss (ablation) | Med-MICN with neural-symbolic loss only · DDI | accuracy: 78.79 -> 81.82 | Helped |
| add neural-symbolic loss (ablation) | Med-MICN with concept loss only · DDI | accuracy: 79.55 -> 81.82 | Helped |
| add concept loss (ablation) | Med-MICN with neural-symbolic loss only · Chest X-Ray | accuracy: 70.03 -> 78.37 | Helped |
| add neural-symbolic loss (ablation) | Med-MICN with concept loss only · Chest X-Ray | accuracy: 72.28 -> 78.37 | Helped |
| add concept loss (ablation) | Med-MICN with neural-symbolic loss only · Fitzpatrick17k | accuracy: 80.79 -> 82.76 | Helped |
| add neural-symbolic loss (ablation) | Med-MICN with concept loss only · Fitzpatrick17k | accuracy: 79.80 -> 82.76 | Helped |
Surprising findings
- Relying solely on the neural-symbolic loss does not notably improve performance over the baseline in the ablation.
Limitations
- Neural symbolic method may need acceleration for large sample sizes.
- Evaluated only on small medical datasets.
Experimental setup
Datasets: COVID-CT DDI Chest X-Ray Fitzpatrick17k
Benchmarks: COVID-CT DDI Chest X-Ray Fitzpatrick17k
Metrics: accuracy precision recall F1 AUC
Baselines: ResNet50 VGG19 DenseNet169 SSSD-COVID Label-free CBM DCR
Builds on: Concept Bottleneck Model Label-free CBM Deep Concept Reasoner BioViL
Closest papers by meaning
- MICA: Towards Explainable Skin Lesion Diagnosis via Multi-Level Image-Concept Alignment (AAAI 2024 · similarity 0.87)
- Interactive Medical Image Analysis with Concept-based Similarity Reasoning (CVPR 2025 · similarity 0.85)
- Interpretable Bilingual Multimodal Large Language Model for Diverse Biomedical Tasks (ICLR 2025 · similarity 0.83)
- SI-MIL: Taming Deep MIL for Self-Interpretability in Gigapixel Histopathology (CVPR 2024 · similarity 0.82)
- Med-CMR: A Fine-Grained Benchmark Integrating Visual Evidence and Clinical Logic for Medical Complex Multimodal Reasoning (CVPR 2026 · similarity 0.82)
- ParseCaps: An Interpretable Parsing Capsule Network for Medical Image Diagnosis (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.