How can medical imaging models discover rare diseases in open-world settings?

Three discovery settings replace one flat unknown class

The paper defines Generalized Biomedicine Discovery (GBD) as a benchmark and problem setting in which a model is trained on labeled known classes and then must assign unlabeled images to either a known class or a newly discovered novel class. The unlabeled set mixes known and novel classes, and its labels are used only for evaluation. This differs from open-set recognition, which the paper says lumps heterogeneous unseen classes into a single "unknown" category, and from novel category discovery, which assumes unlabeled data contain only new classes.

GBD splits that single problem into three clinically grounded regimes. Long-tail rare disease discovery labels only the head classes and asks the model to find tail diseases. Normal-to-abnormal discovery asks it to surface subtle lesions that normal anatomy dominates. Within-taxonomy discovery asks it to find new sibling subtypes inside an existing diagnostic family. The benchmark spans endoscopy, pathology, microscopy, and dermatology, so the claim is about these biomedical imaging structures rather than natural-image label spaces.

The head-tail split defines what counts as rare

In the long-tail setting, the paper orders leaf-level diagnostic categories by empirical frequency n(c) in the training pool and labels only a head fraction controlled by a parameter alpha. With alpha set to 0.5 in general use, the known classes are the smallest set of classes covering half the training samples, and the novel classes are the remaining tail. Ties in frequency are broken lexicographically.

This construction matters for reading any rare-disease number: "novel" here means excluded from the labeled head by that frequency rule, not a clinical rarity threshold. The paper notes that classes explicitly identified as rare in medical data are handled separately, so the alpha = 0.5 default should not be read as a universal definition of rare disease.

Dominant anatomy suppresses weak novelty cues

The paper's central diagnosis is visual rather than taxonomic. Recurring known structure, such as widespread normal anatomy, dominates the global embedding, so weak cues for rare classes, abnormalities, or new sibling subtypes are underrepresented. The authors motivate SCAN as a way to expose deviations relative to each image's dominant content and then integrate them without indiscriminately rewriting established structure.

SCAN is a lightweight plug-in that transforms the class token and patch tokens from a vision backbone while leaving the host discovery objective unchanged. It proceeds in three stages named after a clinical progression: forming a predictive prior, attending to deviations, and accommodating informative evidence.

Filtering, salience, and accommodation in sequence

In predictive suppression, SCAN routes patch tokens into a small set of learnable evidence slots to summarize recurring patterns, averages those slots, and fuses them with the class token into a sample-specific dominant anchor direction. It then removes each patch token's component along that anchor, keeping the residual in the orthogonal complement. The projection uses a stop-gradient on the anchor and requires neither normal labels nor anomaly templates.

In surprise-evoked salience, the residual norm measures how poorly the dominant anchor explains a token. A within-image softmax over residual energies turns those norms into salience weights, and residuals are re-aggregated using both their original slot assignments and their surprise weights. The paper describes this as separating what recurring pattern a token belongs to from how unexpectedly it departs, and it broadcasts the excited evidence back to tokens through a small MLP.

In complementary accommodation, SCAN maps excited evidence into a global novelty proposal using first- and second-order readouts, then removes the proposal's component aligned with the dominant anchor. Pooled token evidence is projected into the same complementary subspace, so the update adds novelty with reduced overlap with dominant content.

Discovery improves while known knowledge is generally kept

The paper reports that SCAN improves novel concept discovery while generally preserving established clinical knowledge, and that it plugs into existing architectures to better navigate the known-unknown trade-off. Those two outcomes are distinct measures: finding novel classes and retaining known-class performance are evaluated separately, and the paper does not claim that one implies the other.

The discussion is more specific than the abstract. It describes overall improvements in novel-disease discovery while revealing some dataset- and baseline-specific trade-offs in preserving established knowledge. Readers should therefore treat preservation as a general tendency reported across experiments, not as a guarantee for every dataset or backbone.

The evidence covers four imaging domains, not deployment

The benchmark covers endoscopy, pathology, microscopy, and dermatology, and the three settings are evaluated as constructed discovery tasks with labels withheld from the unlabeled pool. The paper's claim is that GBD provides a clinically grounded, taxonomy-aware testbed that reframes biomedical imaging as a measurable concept-formation problem.

That framing is the boundary. The reported results concern novel concept discovery and preservation of known clinical knowledge on these benchmark settings; the paper does not report prospective clinical deployment, and the dataset- and baseline-specific trade-offs it acknowledges mean the balance between discovery and retention can shift with the evaluation setup.

Watch the head fraction and the retention baseline

For readers comparing discovery systems on this benchmark, two paper variables carry most of the interpretive weight. The first is the head fraction alpha that defines which diseases are labeled known and which are novel: changing it changes both the rarity of the discovery target and the amount of supervision. The second is the baseline used when judging preservation of known knowledge, because the paper reports dataset- and baseline-specific trade-offs rather than a uniform retention gain.

The paper's own priority is the known-unknown trade-off under realistic biomedical structure: long-tail rarity, normal anatomy dominating subtle lesions, and hierarchical taxonomies. SCAN's contribution is a plug-in that targets the visual suppression behind that trade-off, and the appropriate reading is that it improves novel discovery on these settings while retention must be checked against the specific dataset and baseline in question.

Sources used in this answer