Research map / Probabilistic and scientific modeling / Molecules and proteins
Seeing Unseen: Discover Novel Biomedical Concepts via Geometry-Constrained Probabilistic Modeling
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 a geometry-constrained probabilistic framework (marginal vMF embeddings on a hypersphere with uniform proxies, boundness and uniformity constraints) plus a spectral eigengap method to estimate novel class count for generalized novel class discovery in biomedical images.
Problem
Discovering novel biomedical concepts/classes from unlabeled data under non-i.i.d. distribution shift between base and unlabeled sets and severe class imbalance, while the number of novel classes is unknown.
Approach
Models each instance embedding as a marginal von Mises-Fisher distribution on a hypersphere, anchors base classes to energy-minimized uniform proxies, pushes unlabeled open-set instances away by angular margin, and structures unlabeled data via proxy-based semantic consensus; estimates class number with eigengap of a graph Laplacian.
Key results
- On pneumonia infectious organisms, the method reaches 78.51% Acc-all and 66.67% Acc-novel, versus 72.73% and 60.32% for the strongest baseline DCCL.
- On skin cancer lesions, Acc-novel is 41.67% and F1-novel 41.39%, while on diabetic retinopathy it achieves 60.41% Acc-all versus 54.19% for BYOP (the strongest baseline in that column).
- The spectral class-number estimator takes 0.121s on the pneumonia benchmark, more than 10x faster than the recursive clustering of GCD (2.854s).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use geometry-constrained probabilistic modeling instead of the compared methods | Ours · Pneumonia infectious organisms | F1-novel: 39.79 -> 49.18 | Helped |
| Use geometry-constrained probabilistic modeling instead of the compared methods | Ours · Cell nuclei | F1-novel: 36.57 -> 43.27 | Helped |
| Use geometry-constrained probabilistic modeling instead of the compared methods | Ours · Skin cancer lesions | F1-novel: 23.04 -> 27.12 | Helped |
| Use geometry-constrained probabilistic modeling instead of the compared methods | Ours · Diabetic retinopathy severity | F1-novel: 27.51 -> 34.22 | Helped |
| Add open-space dispersion and structuring (ablation) | Ours · Skin cancer lesions | inter-class angular separation | Helped |
| Add uniform proxies, base-space bounding, open-space dispersion, and structuring (ablation) | Ours · Pneumonia infectious organisms and cell nuclei | F1-novel | Helped |
Surprising findings
- Solely imposing base-space bounding and uniform proxies is insufficient; the full open-space dispersion and structuring components are needed to recover novel-class F1 (ablation on pneumonia F1-novel rises from 23.53 to 49.18).
Limitations
- Evaluated only on image-based biomedical benchmarks with a single ViT-B/16 backbone.
- Novel-class accuracy remains modest in several tasks (e.g., about 41.7% Acc-novel on skin lesions).
Experimental setup
Datasets: pneumonia infectious organism datasets cell nuclei datasets skin lesion datasets (ISIC) diabetic retinopathy datasets
Benchmarks: pneumonia infectious organism discovery cell nuclei discovery skin cancer lesion discovery diabetic retinopathy severity discovery
Metrics: Acc-all Acc-known Acc-novel F1-all F1-known F1-novel
Baselines: RankStats+ UNO+ GCD DCCL SLF BYOP
Builds on: Generalized Category Discovery (GCD) DINO ViT-B/16 SLF BYOP
Closest papers by meaning
- Causal Representation Learning from Multimodal Biomedical Observations (ICLR 2025 · similarity 0.83)
- FIND: A Framework for Discovering Formulas in Data (AAAI 2025 · similarity 0.82)
- Towards scientific discovery with dictionary learning: Extracting biological concepts from microscopy foundation models (ICML 2025 · similarity 0.82)
- Nonparametric Identification of Latent Concepts (ICML 2025 · similarity 0.82)
- Deep Taxonomic Networks for Unsupervised Hierarchical Prototype Discovery (NeurIPS 2025 · similarity 0.82)
- Joint Class-level and Instance-level Relationship Modeling for Novel Class Discovery (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.