Research map / Label-efficient and robust learning / Knowledge distillation
SurvMILKD: A Weakly Supervised Survival Analysis Framework for Multi-Teacher Knowledge Distillation using Pathology Foundation Models
Mayur Mallya, Ali Khajegili Mirabadi, Hossein Farahani, Ali Bashashati
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 |
|---|---|---|---|
| distill four teacher FMs into UNI student with risk-aware loss | UNI + ABMIL · TCGA COAD | C-index (%): 63.02 -> 74.59 | Improved |
| distill into UNI vs 5-FM ensemble on KIRC | SurvMILKD · TCGA KIRC | C-index (%): 71.11 (ensemble) vs 70.96 | Hurt |
| SurvMILKD vs ensemble across datasets | SurvMILKD · six TCGA + two CPTAC3 | C-index (%): e.g. COAD 66.73 -> 74.59; KIRC 71.11 -> 70.96 | Mixed |
| SurvMILKD gives inference cost of a single FM | SurvMILKD · Inference params | M params: 3512.6 -> 303.9 | Improved |
| add risk-aware KD loss to Cox loss (ablation) | SurvMILKD · TCGA COAD | C-index (%): 63.02 -> 71.08 | Improved |
| add risk-aware KD loss to Cox loss (ablation) | SurvMILKD · TCGA KIRC | C-index (%): 69.70 -> 68.51 | Hurt |
| add feature MSE loss on top of KD (ablation) | SurvMILKD · TCGA COAD | C-index (%): 71.08 -> 73.25 | Improved |
| add Cox loss on reconstructed teacher features (ablation) | SurvMILKD · TCGA UCEC | C-index (%): 64.89 -> 67.38 | Improved |
| increase number of teachers from 1 to 4 (ablation) | SurvMILKD · TCGA BRCA | C-index (%): 63.08 -> 68.32 | Improved |
| add more teachers beyond 2-3 (ablation) | SurvMILKD · TCGA KIRC | C-index (%): 71.56 (Vir + CH) vs 70.96 (4 teachers) | No clear effect |
| use a weak teacher (H-Optimus-0 below UNI on COAD) alone (ablation) | SurvMILKD · TCGA COAD | C-index (%): 63.02 -> 69.82 | Improved |
| use other FMs as student (e.g. H-Optimus-0 on UCEC) (ablation) | SurvMILKD · TCGA UCEC | C-index | Hurt |
Closest papers by meaning
- Interpretable Vision-Language Survival Analysis with Ordinal Inductive Bias for Computational Pathology (ICLR 2025 · similarity 0.88)
- Revisiting End-to-End Learning with Slide-level Supervision in Computational Pathology (NeurIPS 2025 · similarity 0.87)
- FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot Whole Slide Image Classification (CVPR 2025 · similarity 0.87)
- Leveraging Tumor Heterogeneity: Heterogeneous Graph Representation Learning for Cancer Survival Prediction in Whole Slide Images (NeurIPS 2024 · similarity 0.86)
- Modeling Dense Multimodal Interactions Between Biological Pathways and Histology for Survival Prediction (CVPR 2024 · similarity 0.86)
- Scaling Multi-Teacher Distillation for Digital Pathology (NeurIPS 2026 · similarity 0.86)
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.
