Generalized Biomedicine Discovery¶
Conference: ECCV 2026
Paper: ECCV Official
Code: https://github.com/lytang63/generalized-biomedicine-discovery
Area: Medical Imaging
Keywords: Generalized Biomedicine Discovery, Generalized Category Discovery, Cognitive Perception, Predictive Suppression, Surprise Salience
TL;DR¶
Addressing the suppression of subtle lesions and rare diseases by dominant normal anatomy in clinical open-world environments, this paper introduces the Generalized Biomedicine Discovery (GBD) benchmark and SCAN, a cognition-inspired plug-and-play vision layer that filters predictable patterns via predictive suppression, excites anomalous salience via surprise energy, and performs complementary accommodation to discover novel concepts while preserving established clinical knowledge.
Background & Motivation¶
Deploying medical imaging systems in real-world clinical environments requires operating under open-world conditions far beyond closed-world assumptions. Real clinical data consistently manifest three defining open-world signatures: long-tailed class distributions where clinically critical rare diseases are sparsely observed; visual dominance of normal anatomy where expansive healthy tissue overwhelms localized, subtle lesions; and hierarchical diagnostic taxonomies where novel concepts emerge as fine-grained sibling subtypes within established disease families. However, standard open-set recognition (OSR) collapses all unseen categories into a single uninformative "unknown" label, while conventional novel category discovery (NCD) and generalized category discovery (GCD) predominantly assume flat, balanced concept spaces in natural images, leaving these authentic biomedical challenges unaddressed.
The core tension stems from how clinicians versus deep visual models perceive disease: expert clinicians diagnostic reasoning relies on an established mental model of normal anatomy to detect subtle unexpected deviations and situate new subtypes into diagnostic families. In contrast, deep vision backbones optimize global representation learning across images, causing high-frequency normal anatomical structures to form a dominant visual manifold. This dominant manifold systematically suppresses residual evidence of subtle abnormalities and rare diseases, preventing standard GCD models from reliably discriminating novel clinical concepts in mixed unlabeled data.
To resolve this limitation, the goal is to formalize realistic biomedical discovery paradigms and decouple dominant visual suppression from novel concept formation. Core idea: inspired by clinical diagnostic cognition, SCAN instantiates a three-stage perceptual progressionβpredictive suppression of dominant patterns, surprise-evoked salience amplification, and complementary feature accommodationβto isolate and assimilate subtle novel evidence in the orthogonal complement of the dominant manifold.
Method¶
Overall Architecture¶
Given an input biomedical image, a visual encoder (such as ViT-B/16) extracts a global class token \(\mathbf{c} \in \mathbb{R}^D\) and \(N\) spatial patch tokens \(\mathbf{X} = [\mathbf{x}_1, \dots, \mathbf{x}_N]^\top \in \mathbb{R}^{N \times D}\). SCAN functions as a lightweight, plug-and-play layer that transforms features \((\mathbf{c}, \mathbf{X}) \mapsto \tilde{\mathbf{c}}\) while leaving the host GCD model's clustering loss and training protocol completely unchanged. The end-to-end flow proceeds across three stages: first, patch tokens are softly routed to learnable evidence slots to establish an image-specific dominant anchor, geometrically filtering predictable anatomical components; next, patch-level residual energy is quantified as Bayesian surprise to compute salience weights, aggregating excited evidence of novelty and broadcasting it back to patch tokens; finally, the novelty proposal is projected into the orthogonal complement of the dominant anchor and modulated by a non-parametric surprise gate before being additively integrated into the global class representation.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Input Image & ViT Features<br/>Class Token c and Patch Tokens X"] --> B["Predictive Suppression & Geometric Filtering<br/>Route evidence slots, form dominant anchor u, project out predictable norm"]
B --> C["Surprise-Evoked Salience & Evidence Excitation<br/>Residual norm yields surprise energy si and salience weights wi"]
C --> D["Orthogonal Complement Projection & Gated Accommodation<br/>Project novelty proposal to orthogonal subspace, inject via gate Ξ³"]
D --> E["Updated Representation c_tilde<br/>Enter downstream GCD clustering and classification"]
Key Designs¶
1. Predictive Suppression & Geometric Filtering: Constructing dominant anchor and filtering expected norms Targeting the problem where dominant normal anatomy and frequent head classes overwhelm subtle novel signals, and drawing inspiration from visual predictive coding where feedback connections actively suppress predictable sensory inputs, SCAN avoids relying on static normal prototypes. Instead, it introduces \(E\) learnable evidence prototypes \(\mathbf{P} \in \mathbb{R}^{E \times D}\) as an image-adaptive codebook of dominant visual patterns. Soft routing weights with temperature \(\tau_r\) are computed as \(A_{ij} = \frac{\exp(\langle \mathbf{x}_i, \mathbf{p}_j \rangle / \tau_r)}{\sum_{k=1}^E \exp(\langle \mathbf{x}_i, \mathbf{p}_k \rangle / \tau_r)}\) and normalized over tokens into \(\tilde{A}_{ij} = A_{ij} / \sum_{k=1}^N A_{kj}\). These yield \(E\) evidence slot summaries \(\mathbf{e}_j = \sum_{i=1}^N \tilde{A}_{ij} \mathbf{x}_i\) and their mean \(\bar{\mathbf{e}} = \frac{1}{E} \sum_{j=1}^E \mathbf{e}_j\). Fusing global and local slot evidence forms an image-specific dominant anchor direction: $\(\mathbf{u} = \frac{\mathbf{c} + \bar{\mathbf{e}}}{\|\mathbf{c} + \bar{\mathbf{e}}\|}\)$ Through geometric orthogonal projection, predictable components along the anchor direction are actively subtracted, producing residual patch tokens \(\mathbf{R} = [\mathbf{r}_1, \dots, \mathbf{r}_N]^\top\): $\(\mathbf{r}_i = \mathbf{x}_i - (\mathbf{x}_i^\top \text{sg}(\mathbf{u})) \text{sg}(\mathbf{u})\)$ where \(\text{sg}(\cdot)\) denotes stop-gradient to stabilize the anchor. This operation filters expected anatomical baselines without requiring explicit normal annotations or lesion templates.
2. Surprise-Evoked Salience & Evidence Excitation: Quantifying unexpected deviations via residual energy After filtering out dominant patterns, the residual vector magnitude directly reflects local discrepancy from predictable anatomy. Drawing upon Bayesian surprise theory, SCAN defines patch-wise surprise energy as the residual norm \(s_i = \|\mathbf{r}_i\|_2\) and converts it into a normalized within-image salience distribution \(w_i = \frac{\exp(s_i / \tau_s)}{\sum_{k=1}^N \exp(s_k / \tau_s)}\), where \(\tau_s\) controls concentration. Resynthesizing residuals via their slot assignments and surprise weights yields excited novelty evidence: $\(\mathbf{e}_j^+ = \sum_{i=1}^N \tilde{A}_{ij} w_i \mathbf{r}_i\)$ To refine local representations, the excited evidence matrix \(\mathbf{E}^+ = [\mathbf{e}_1^+, \dots, \mathbf{e}_E^+]^\top\) is broadcast back to patch tokens using a lightweight MLP \(\phi(\cdot)\): \(\mathbf{X}^{out} = \mathbf{X} + \phi(\tilde{\mathbf{A}} \mathbf{E}^+)\). This mechanism decouples pattern semantics from deviation magnitude, amplifying subtle pathological anomalies.
3. Orthogonal Complement Projection & Gated Accommodation: Non-interfering assimilation of novel concepts Naively injecting isolated deviations into global features risks disrupting established classification boundaries. Grounded in Complementary Learning Systems (CLS) theory, SCAN maps excited evidence into a global novelty proposal incorporating first-order signed mean shift and second-order dimension-wise energy: \(\mathbf{\Delta} = \mathbf{W}_1 \left(\frac{1}{E}\sum_{j=1}^E \mathbf{e}_j^+\right) + \mathbf{W}_2 \left(\left(\frac{1}{E}\sum_{j=1}^E \mathbf{e}_j^+\right) \odot \left(\frac{1}{E}\sum_{j=1}^E \mathbf{e}_j^+\right)\right)\), with learnable weights \(\mathbf{W}_1, \mathbf{W}_2 \in \mathbb{R}^{D \times D}\). To prevent direct interference with known patterns, both \(\mathbf{\Delta}\) and pooled token features \(\mathbf{x}_{pool} = \text{MeanPool}(\mathbf{X}^{out})\) are projected into the orthogonal complement of the anchor \(\mathbf{u}\): $\(\mathbf{\Delta}_\perp = \mathbf{\Delta} - (\mathbf{\Delta}^\top \mathbf{u})\mathbf{u}, \qquad \mathbf{x}_{pool\perp} = \mathbf{x}_{pool} - (\mathbf{x}_{pool}^\top \mathbf{u})\mathbf{u}\)$ To regulate update scale, a calibration-free cognitive gate \(\gamma\) is formulated using three complementary signals: surprise magnitude \(S = \sum_{i=1}^N w_i s_i\), knownness alignment \(k = \cos(\mathbf{c}, \bar{\mathbf{e}})\), and peakedness \(\kappa = 1 - \frac{\text{Ent}(\mathbf{w})}{\log N}\). Standardizing \(S\) and \(k\) via exponential moving average (EMA) statistics produces \(\hat{S}\) and \(\hat{k}\), yielding the gate: $\(\gamma = \sigma(\hat{S} - \hat{k}) \cdot \text{clip}(\kappa, 0, 1)\)$ The final representation is formed via stable residual integration: \(\tilde{\mathbf{c}} = \mathbf{c} + \gamma \cdot \text{LN}(\mathbf{\Delta}_\perp + \mathbf{x}_{pool\perp})\). When a sample is well explained by dominant evidence, \(\gamma \to 0\) recovers the original representation to protect known classes; when unexpected salient deviations occur, orthogonal novelty is selectively assimilated.
Loss & Training¶
SCAN is integrated directly as a feature refinement block, preserving each host GCD baseline's loss functions and training schedules without modification. The backbone is a ViT-B/16 initialized with self-supervised DINO-V2 weights, trained across 200 epochs on mixed labeled and unlabeled sets. SCAN uses \(E=16\) evidence slots by default across all datasets. At evaluation time, the Hungarian optimal assignment algorithm is solved once over all classes in the unlabeled pool to measure clustering accuracy across All, Old (known), and New (novel) subsets.
Key Experimental Results¶
Main Results¶
The GBD benchmark spans four distinct biomedical domains: GI endoscopy (Gastrovision), multi-organ histopathology (HistoSet-5Γ14), peripheral blood cytology (MLL23), and dermoscopy (Derm12345). The table below details performance across Setting I (Long-tail Rare Disease Discovery) and Setting II (Normal-to-Abnormal Discovery), reporting clustering accuracy (ACC %):
| Setting & Method | Gastrovision (All / Old / New) | HistoSet (All / Old / New) | MLL23 (All / Old / New) | Derm12345 (All / Old / New) | Average (All / Old / New) |
|---|---|---|---|---|---|
| Setting I: Long-tail Discovery | |||||
| SEALβ (hierarchical supervision) | 56.8 / 51.2 / 60.7 | 81.2 / 87.1 / 77.9 | 65.4 / 67.6 / 54.1 | 56.5 / 61.0 / 10.6 | 65.0 / 66.7 / 50.8 |
| SimGCD | 37.8 / 40.6 / 31.8 | 74.4 / 83.8 / 71.8 | 56.9 / 52.9 / 67.0 | 25.4 / 27.9 / 13.1 | 48.6 / 51.3 / 45.9 |
| SimGCD + SCAN | 39.7 / 43.3 / 31.9 | 80.0 / 82.1 / 79.4 | 64.3 / 63.1 / 67.2 | 25.9 / 28.2 / 14.2 | 52.5 / 54.2 / 48.1 |
| LegoGCD | 39.8 / 48.9 / 26.1 | 72.5 / 76.0 / 71.6 | 70.8 / 63.4 / 89.7 | 26.3 / 28.5 / 15.7 | 52.3 / 54.2 / 50.8 |
| LegoGCD + SCAN | 41.3 / 49.4 / 29.2 | 72.5 / 77.5 / 71.2 | 73.2 / 67.2 / 88.8 | 26.7 / 29.0 / 14.8 | 53.4 / 55.8 / 51.0 |
| SelEx | 50.1 / 52.0 / 47.4 | 76.4 / 97.5 / 70.5 | 66.4 / 68.3 / 61.3 | 28.9 / 29.1 / 27.9 | 55.5 / 61.7 / 51.8 |
| SelEx + SCAN (Ours) | 71.2 / 83.4 / 53.0 | 76.3 / 99.9 / 69.7 | 80.7 / 89.3 / 58.6 | 35.6 / 35.6 / 35.6 | 66.0 / 77.1 / 54.2 |
| Setting II: Normal-to-Abnormal | |||||
| SEALβ (hierarchical supervision) | 62.6 / 64.2 / 61.0 | 73.5 / 98.7 / 48.2 | 69.2 / 70.9 / 60.1 | 59.0 / 61.7 / 5.5 | 66.1 / 73.8 / 43.7 |
| SimGCD | 39.7 / 42.9 / 25.6 | 61.1 / 94.1 / 44.5 | 57.9 / 55.9 / 63.4 | 26.6 / 27.2 / 19.9 | 46.3 / 55.0 / 38.3 |
| SimGCD + SCAN | 40.1 / 44.0 / 22.6 | 62.0 / 93.6 / 46.1 | 58.3 / 57.5 / 60.6 | 26.4 / 26.8 / 21.6 | 46.7 / 55.5 / 37.7 |
| LegoGCD | 49.3 / 53.7 / 24.9 | 62.0 / 97.5 / 44.2 | 68.5 / 72.1 / 58.6 | 28.5 / 30.8 / 5.5 | 52.1 / 63.5 / 33.3 |
| LegoGCD + SCAN | 50.5 / 55.5 / 23.1 | 63.9 / 97.4 / 47.1 | 69.4 / 73.7 / 57.7 | 29.0 / 29.7 / 21.6 | 53.2 / 64.1 / 37.4 |
| SelEx | 50.3 / 51.9 / 41.7 | 64.3 / 97.7 / 47.5 | 58.5 / 58.0 / 59.8 | 29.3 / 28.2 / 40.7 | 50.6 / 58.9 / 47.4 |
| SelEx + SCAN (Ours) | 70.1 / 86.0 / 46.3 | 68.6 / 97.2 / 60.7 | 87.1 / 89.1 / 81.7 | 33.6 / 33.3 / 36.4 | 64.8 / 76.4 / 56.3 |
Ablation & Overhead Study¶
The following table details computational complexity and performance across Setting III (Within-Taxonomy Discovery) across different host models:
| Configuration / Host Model | Params (M) | Training Time (s) | Inference Time (s) | Setting III Avg All ACC (%) | Setting III Avg New ACC (%) |
|---|---|---|---|---|---|
| SimGCD Baseline | 92.10 | 27.87 | 10.61 | 51.2 | 48.9 |
| SimGCD + SCAN | 99.68 (+7.6M) | 30.12 (+8%) | 11.32 (+6%) | 50.7 (-0.5) | 48.1 (-0.8) |
| LegoGCD Baseline | - | - | - | 47.5 | 43.7 |
| LegoGCD + SCAN | - | - | - | 50.5 (+3.0) | 46.8 (+3.1) |
| SelEx Baseline | - | - | - | 53.7 | 49.8 |
| SelEx + SCAN (Full Model) | - | - | - | 66.0 (+12.3) | 53.1 (+3.3) |
Key Findings¶
- Amplified gains with fine-grained self-expertise: SCAN demonstrates maximum synergy when combined with SelEx. In Setting I (long-tail), SelEx + SCAN improves average All accuracy by +10.5% (55.5% to 66.0%); in Setting II (normal-to-abnormal), average All accuracy increases by +14.2% (50.6% to 64.8%) with a +8.8% gain on abnormal New classes; in Setting III (within-taxonomy), All accuracy rises by +12.3% (53.7% to 66.0%).
- Outperforming hierarchically supervised models: While SEAL requires ground-truth taxonomic hierarchy annotations during training, SelEx + SCAN outperforms SEAL in several benchmarks without explicit hierarchical labels (e.g., Setting I average All 66.0% vs. 65.0%; Setting II MLL23 All 87.1% vs. 69.2%).
- Minimal computational overhead: SCAN adds only 7.6M parameters (+8.2%), increasing training runtime by 8% and inference latency by 6%, maintaining practical clinical efficiency.
Highlights & Insights¶
- Bridging cognitive psychology and clinical visual manifolds: Successfully operationalizes predictive coding, Bayesian surprise, and complementary learning systems into concrete linear-algebraic operators (orthogonal geometric filtering, residual energy weighting, and complementary projection), eliminating dominant normal tissue masking.
- Calibration-free dynamic cognitive gating: By tracking surprise magnitude, knownness alignment, and entropy-based concentration through running EMA moments, the gate \(\gamma\) dynamically balances novel discovery against catastrophic forgetting across heterogeneous modalities.
- Transferable orthogonal residual projection: Projecting updates strictly onto the orthogonal complement of dominant feature directions offers a general-purpose blueprint for unlearning dominant biases in long-tail learning and open-set recognition.
Limitations & Future Work¶
- Host baseline sensitivity: While SCAN yields massive gains on SelEx, its improvements on SimGCD are marginal or show slight trade-offs on specific subsets, indicating that orthogonal residual features require fine-grained clustering objectives to fully exploit.
- Fixed evidence slot budget: Using \(E=16\) slots universally across all imaging modalities preserves fairness, but multi-scale biomedical images (e.g., single-cell cytology vs. endoscopic macroscopic fields) likely benefit from adaptive slot sizing.
- Future directions: Integrating SCAN into multimodal biomedical vision-language models (e.g., BiomedCLIP) and exploring dynamically scaled slot routing conditioned on image resolution.
Related Work & Insights¶
- vs SEAL: SEAL relies on explicit multi-level taxonomic trees for hierarchical discovery, whereas SCAN tackles the upstream feature suppression problem without tree supervision, achieving superior novel class discovery across multiple benchmarks.
- vs LegoGCD / SimGCD: Standard GCD methods rely on global contrastive representations, implicitly assuming clean class separability; SCAN acts as a foundational cognitive layer that extracts faint residual evidence prior to category clustering.
Rating¶
- Novelty: βββββ Formulates the clinically grounded GBD benchmark and designs a three-stage cognitive perceptual layer operating in the orthogonal complement space.
- Experimental Thoroughness: βββββ Thoroughly evaluated across 4 biomedical modalities, 3 clinically motivated discovery settings, and 3 distinct GCD host baselines.
- Writing Quality: βββββ Rigorous conceptual framing, clear mathematical formulations, and insightful qualitative and dynamical analyses.
- Value: βββββ Offers a principled paradigm for discovering emerging diseases and rare subtypes in clinical open-world archives.