Source-linked AI summary
Generalized Biomedicine Discovery
Luyao Tang, Yingkai Yang, Hanqi Chen, Jiewei Zheng, Chaoqi Chen, Cheng Chen
TL;DR
Clinical discovery must address rare diseases, subtle lesions, and hierarchical taxonomies that flat, balanced open-world paradigms leave unresolved. GBD benchmarks these settings, and SCAN applies predictive suppression, surprise-evoked salience, and complementary accommodation; experiments show overall improvements in novel-disease discovery.
Problem
Medical imaging discovery faces long-tailed rare diseases, subtle lesions dominated by normal anatomy, and hierarchical taxonomies, while most open-world paradigms assume flat, relatively balanced concept spaces.
Source: §1 · p. 1Method
GBD benchmarks long-tail rare disease, normal-to-abnormal, and within-taxonomy discovery; SCAN uses predictive suppression, surprise-evoked salience, and complementary accommodation.
Source: §1 · p. 2, §6 · p. 14Results
SCAN generally improves conventional GCD methods on GBD, including +14.2% average All accuracy for SelEx on normal-to-abnormal discovery.
Source: §1 · p. 3, §5.2 · p. 12Takeaways & Limitations
GBD targets upstream clinical concept formation by organizing unlabeled archives into cohorts experts can verify, name, merge, or split.
Source: §5.4 · p. 14Some baseline- and subset-specific trade-offs remain.
Source: §1 · p. 3
Abstract
from arXiv · showhide
In real-world clinical practice, medical images face open-world shifts: (i) long-tailed rare diseases, (ii) subtle lesions dominated by normal anatomy, and (iii) hierarchical taxonomies. Yet most open-world paradigms assume flat, balanced label spaces, leaving these biomedical demands unresolved. We introduce Generalized Biomedicine Discovery (GBD) and a unified benchmark spanning long-tail, anomaly, and taxonomy-aware discovery. Our key insight is that dominant known patterns form a visual manifold that masks subtle novelty. Inspired by expert diagnosis, we propose SCAN (Surprise-evoked Complementary AccommodatioN), which follows a cognition-inspired perceptual progression: it applies predictive suppression to filter expected norms, triggers surprise-evoked salience to highlight unexpected deviations, and performs complementary accommodation to integrate these shifts into global representations. Extensive experiments show that SCAN improves novel concept discovery while generally preserving established clinical knowledge, and it plugs into existing architectures to better navigate the known-unknown trade-off in medical imaging. Code is available at https://github.com/lytang63/generalized-biomedicine-discovery.
1 Introduction
Clinical images present open-world discovery challenges that flat, balanced paradigms do not capture. The paper introduces GBD and SCAN to address these challenges, reporting that SCAN generally improves discovery on GBD.
1 Introduction
- Open-world medical imaging combines rare diseases, subtle lesions masked by normal anatomy, and hierarchical taxonomies, while most existing paradigms assume flat, relatively balanced concept spaces.Methods validated on natural-image GCD often fail to translate to clinical settings.Source: §1 · pp. 1–2
- GBD is a unified medical-imaging benchmark with long-tail rare disease, normal-to-abnormal, and within-taxonomy discovery settings.The settings reflect the challenge of discovering rare conditions, subtle lesions, and new sibling subtypes.Source: §1 · p. 2, Figure 1 · p. 2
- SCAN is a plug-and-play layer that suppresses anticipated patterns, amplifies unexpected localized deviations, and integrates novelty while preserving established structure.It operates on a vision backbone’s class and patch tokens.Source: §1 · p. 3
- SCAN generally improves conventional GCD methods on GBD, with particularly strong SelEx gains for subtle anomalies and long-tailed rare diseases, though some baseline- and subset-specific trade-offs remain.The introduction summarizes the experiments as improving discovery while retaining trade-offs.Source: §1 · p. 3
2 Related Work
Prior open-world learning methods extend beyond closed-set recognition, but biomedical class imbalance, dominant normal anatomy, and clinical hierarchies remain underexplored. The paper connects these gaps to a unified framework informed by cognitive theories.
Open-world Learning
- Open-set recognition groups unseen classes as one unknown, novel category discovery assumes unlabeled data contains only novel classes, and generalized category discovery handles mixed known and novel classes.Despite these advances, biomedical imbalance, normal anatomy, and clinical hierarchies remain underexplored.Source: §2 · p. 3
Cognitive Sciences in Machine Learning
- Cognitive machine-learning work has used perceptual grouping, inductive reasoning, structured priors, and factorized representations to support robust generalization.The cited work spans several tasks and approaches.Source: §2 · p. 3
- Predictive coding, surprise-driven salience, and complementary learning systems have been explored separately but remain underused for biomedical diagnostic cognition; the paper integrates them into one framework.The framework is presented as addressing biomedical open-world learning.Source: §2 · p. 4
3 Problem Setting
GBD evaluates discovery when unlabeled data mixes known and novel classes, across long-tail, normal-to-abnormal, and within-taxonomy settings. Its partitions model rare classes among common ones, abnormalities beyond normal categories, and unseen siblings within diagnostic families.
3.1 Task Backbone: Generalized Biomedicine Discovery (GBD)
- GBD provides labeled known-class samples and unlabeled samples from both known and novel classes; the objective is to assign each unlabeled sample to a known or newly discovered class.The labeled subset contains labels only for known classes.Source: §3.1 · p. 4
3.2 Setting I: Long-tail Rare Disease Discovery
- The long-tail setting partitions leaf-level diagnostic categories by empirical frequency so common conditions are known and rare tail diseases must be discovered.The usual head–tail split uses α = 0.5, except for classes explicitly identified as rare in medical references.Source: §3.2 · p. 4
3.3 Setting II: Normal-to-Abnormal Discovery
- Normal-to-abnormal discovery labels normal subclasses as known and pathological categories as novel, testing generalization beyond a normal manifold.The normal categories may include distinct anatomical landmarks, while the abnormal categories include multiple pathologies.Source: §3.3 · p. 4
3.4 Setting III: Within-Taxonomy Discovery
- Within-taxonomy discovery partitions sibling leaf classes into known and novel sets, ensuring each eligible parent group has at least one of each.A smaller ρ yields fewer labeled siblings and more novel siblings to discover, testing fine-grained novelty near known diagnostic families.Source: §3.4 · p. 5
4 Method
SCAN targets the shared suppression of rare, abnormal, and fine-grained novel cues by recurring known structure. It transforms class and patch tokens through predictive suppression, surprise-evoked salience, and complementary accommodation while leaving the host discovery objective unchanged.
4.1 Stage I: Predictive Suppression
- Predictive suppression builds an image-conditioned dominant anchor and removes each patch token’s aligned component, preserving remaining geometry without normal labels or anomaly templates.The anchor represents predictable content, such as dominant normal anatomy.Source: §4.1 · pp. 6–7
4.2 Stage II: Surprise-Evoked Salience
- Residual mismatch measures patch-wise surprise; within-image salience weights rank that evidence, while slot semantics preserve which recurring pattern each token belongs to.Accommodation uses excited evidence re-aggregated from residuals with both slot assignments and surprise weights.Source: §4.2 · p. 7
4.3 Stage III: Complementary Accommodation
- Complementary accommodation combines first- and second-order novelty evidence, gates the update by surprise, knownness, and salience concentration, and biases it away from the dominant direction.The identity path recovers the original representation when the gate is small; the updated representation enters the unchanged GCD pipeline. The anchor represents head patterns, normal anatomy, or parent morphology across GBD’s three settings, without setting-specific supervision.Source: §4.3 · pp. 8–9
5 Experiments
Experiments evaluate GBD across four imaging domains and three clinically grounded discovery settings, then assess SCAN’s performance, dynamics, qualitative behavior, and computational overhead.
5.1 Experimental Setup
- The setup spans four biomedical imaging datasets and evaluates known and novel subsets after one global Hungarian assignment, largely retaining each host baseline’s original training protocol.SCAN is tested with representative GCD baselines for forgetting mitigation and hierarchy-aware learning.Source: §5.1 · pp. 9–10
5.2 Main Results
- SelEx+SCAN gains average All accuracy by +10.5% in long-tail, +14.2% in normal-to-abnormal, and +12.3% in within-taxonomy discovery.The results also report subset-specific regressions in some settings.Source: Table 1 · p. 10, §5.2 · pp. 11–12, Table 2 · p. 11
5.3 Mechanism and Qualitative Analysis
- On Gastrovision, SCAN keeps old accuracy comparable to SelEx while novel accuracy rises; increasing surprise signals accompany the gains, and knownness slightly drops.Around 40 epochs, SelEx shows a performance drop while SCAN enters a second growth phase.Source: §5.3 · p. 13, Figure 3 · p. 13
- On MLL23, SCAN’s embedding visualization shows tighter neighborhoods and clearer class margins; evidence slots recur around punctate textures and diffuse staining patterns.These qualitative patterns suggest that learned slots may provide reusable morphological primitives across samples.Source: §5.3 · p. 13, Figure 4 · p. 14
- SCAN adds a small number of parameters and several linear layers, with limited measured increases in training and inference time.The authors describe the resulting computational overhead as modest for integration into existing discovery pipelines.Source: §5.3 · p. 14, Figure 4 · p. 14
5.4 Discussion
- GBD targets upstream clinical concept formation from unlabeled archives, not leaderboard comparison with text-supervised multimodal models.The intended workflow lets experts verify, name, merge, or split cohorts before integration into guidelines and curated datasets.Source: §5.4 · p. 14
6 Conclusion
The conclusion presents GBD as a clinically grounded benchmark for three open-world discovery settings and SCAN as a cognition-inspired approach to weak novelty cues. Experiments show improved novel-disease discovery alongside dataset- and baseline-specific trade-offs in preserving established knowledge.
6 Conclusion
- GBD establishes a unified benchmark for long-tail, normal-to-abnormal, and within-taxonomy discovery across endoscopy, pathology, microscopy, and dermatology.Source: §6 · p. 14
- SCAN combines predictive suppression, surprise-evoked salience, and complementary accommodation to address dominant known patterns masking weak novelty cues.Source: §6 · p. 14
- Experiments show overall improvements in novel-disease discovery, with dataset- and baseline-specific trade-offs in preserving established knowledge.Source: §6 · p. 14