BioCluster

Section 21

Nearest-neighbor retrieval

For a sample of query images, how often do the k nearest neighbors in feature space share the query's ground-truth label? This reveals morphological coherence independent of any clustering algorithm's choices.

PathMNIST (histopathology)

RepresentationPrecision@1Precision@5Precision@10Queries
PCA on pixels0.5740.5570.531500
Handcrafted color/texture0.2940.2880.286500
HOG0.2320.2030.198500
ResNet50 (ImageNet-supervised)0.8060.7800.741500
DINO ViT-S/160.9460.8920.860500
DINOv2 ViT-S/140.8820.8520.828500

BloodMNIST (blood cell microscopy)

RepresentationPrecision@1Precision@5Precision@10Queries
PCA on pixels0.5800.5200.498500
Handcrafted color/texture0.2660.2580.255500
HOG0.4440.3690.364500
ResNet50 (ImageNet-supervised)0.6040.5440.511500
DINO ViT-S/160.7240.6780.641500
DINOv2 ViT-S/140.6400.5660.536500
Precision@k = fraction of a query's k nearest neighbors (excluding itself) that share its ground-truth label, averaged over a random sample of queries per representation. Labels are used only to evaluate retrieval quality after the fact, never to influence which neighbors are retrieved.