Supervised vs. adversarially pre-trained pseudo-label generators for semi-supervised breast cancer patch classification

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Automated breast cancer classification from whole-slide images (WSIs) is hampered by the scarcity of expert-annotated patches and the domain variability inherent to multi-institutional datasets. We propose and compare two semi-supervised frameworks for patch-level tumor classification under limited annotation. The first, a semi-supervised baseline , trains a supervised model on the available labeled data, uses it to generate confidence-filtered pseudo-labels for the unlabeled corpus, and trains a separate classifier from scratch on the combined data using a joint supervised and pseudo-label objective. The second, an AdCo-augmented framework , replaces the supervised pseudo-label generator with a backbone pre-trained on 347,884 unlabeled CPTAC-BRCA patches using Adversarial Contrastive Learning (AdCo), followed by frozen-backbone linear probing to generate pseudo-labels; a separate model is then trained with independent ImageNet initialization using the same semi-supervised objective. AdCo replaces the conventional fixed memory queue with a bank of learnable adversarial negatives optimized via gradient ascent, compelling the encoder to learn more discriminative histological representations than standard contrastive methods. Both frameworks are evaluated across nine architecture combinations each—18 in total—drawn from three backbones (ResNet-18, DenseNet-121, and ConvNeXt-Small) on TIGER, BRACS, TCGA, and CPTAC-BRCA datasets, providing a systematic benchmark for semi-supervised breast cancer WSI classification and quantifying the benefit of adversarial self-supervised pre-training over purely supervised pseudo-label generation.