Can 1,000 images be enough if the loss knows the feature space?

Contrast-enhanced spectral mammography is used when a mammogram is not enough — dense tissue, suspicious lesion management. The public data amount to about a thousand images.

SimCLR pretraining helps, but not by much at that scale — which left the loss function as the remaining place to find leverage.

After pretraining, every training feature is banked once. For each sample, its nearest same-class and farthest different-class neighbours are found by L2 distance. If a sample's neighbours mostly carry a different label, it is a hard sample and gets more weight in the cross-entropy.

The same neighbours serve as triplet positives and negatives directly, so there is no Siamese network and no online mining — the whole thing is precomputed.

CDD-CESM: 1,003 subtracted CESM images from 326 patients at Cairo University, with pixel-level lesion annotations. ROIs were extracted for benign and malignant lesions, with a fixed 512 × 512 centre crop for normals, split at patient level into 745 / 130 / 116.

Cross-entropy, focal loss and DWNR were compared under both from-scratch and ImageNet-initialized SimCLR pretraining, five runs each.

In the ImageNet + SimCLR track, DWNR lifted F1 from 0.8159 to 0.8413 and AUC from 0.9278 to 0.9350 over cross-entropy, with focal loss landing below both. The gain concentrates on the benign class, where per-class F1 went from 0.6589 to 0.7058.

It also transferred: initializing a whole-image CESM classifier from the DWNR-trained ROI model beat ImageNet initialization by more than four points of F1 on both BI-RADS management and diagnosis.

The result stands: geometry in a self-supervised feature space carries usable training signal in a small-data domain, and the improvement transfers to the whole-image task.

But three quantities here had to be swept per dataset — neighbour count, difficulty weights, triplet margin. That is exactly the limitation the later DITL framework removes by deriving all three from the data.