Research map / Detection and segmentation / Open-set and few-shot detection
Verifying Cancer Segmentation in Vision Transformers via Internal Concepts
Mengmeng Ma, Yunxiang Peng, Tang Li, Lu Lin, Binsheng Zhao, Oguz Akin, Xi Peng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SAE internal concepts instead of output confidence for failure detection | failure detector · Prostate158 zero-shot | AUROC: 50.0 -> 73.6 | Improved |
| use SAE internal concepts instead of MaxProb | failure detector · Prostate158 zero-shot | FPR95: 100.0 -> 64.4 | Improved |
| use SAE concepts instead of output cues on prostate cancer | failure detection and correction · PI-CAI | DSC and F1: +8.7 DSC, +6.2 F1 | Improved |
| use SAE concepts instead of output cues on pancreatic cancer | failure detection and correction · PanTS | DSC and F1: +2.3 DSC, +9.2 F1 | Improved |
| add early-layer concepts only vs all layers (ablation) | SAE concept detector · PI-CAI | F1: 44.5 -> 63.5 | Improved |
| add concepts from all layers instead of deep layers only (ablation) | SAE concept detector · PI-CAI | F1: 57.6 -> 63.5 | Improved |
| add concepts from all layers instead of mid layers only (ablation) | SAE concept detector · PI-CAI | DSC: 48.8 -> 50.1 | Improved |
| use XGBoost instead of logistic regression (ablation) | failure classifier · PI-CAI | F1: 59.2 -> 63.5 | Improved |
| use XGBoost instead of random forest (ablation) | failure classifier · PI-CAI | DSC: 44.6 -> 50.1 | Improved |
| increase dictionary size from 1,536 to 3,072 at sparsity 8 (ablation) | SAE · PI-CAI | F1: 63.5 -> 57.4 | Hurt |
| raise sparsity from 4 to 8 at dictionary size 1,536 (ablation) | SAE · PI-CAI | F1: 60.3 -> 63.5 | Improved |
| raise sparsity from 8 to 16 at dictionary size 1,536 (ablation) | SAE · PI-CAI | F1: 63.5 -> 62.5 | Hurt |
Closest papers by meaning
- "Why Is There a Tumor?": Tell Me the Reason, Show Me the Evidence (ICML 2025 · similarity 0.86)
- Disentangling Human Error from Ground Truth in Segmentation of Medical Images (NeurIPS 2020 · similarity 0.86)
- Make Foundation Models Trustworthy Again: Causal Fine-Adaptation for Medical Image Segmentation (AAAI 2026 · similarity 0.86)
- Med-SegLens: Latent-Level Model Diffing for Interpretable Medical Image Segmentation (ICML 2026 · similarity 0.85)
- Synthesize then Compare: Detecting Failures and Anomalies for Semantic Segmentation (ECCV 2020 · similarity 0.85)
- Interpretable Failure Detection with Human-Level Concepts (AAAI 2025 · similarity 0.84)
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.
