AAC-TTA: Anti-adversarial consistency with test-time adaptation for imbalanced histopathology classification under...

The performance of medical image classification models can degrade at deployment because of two common real-world challenges: domain shift and class imbalance. Domain shift causes whole-slide images acquired on different scanners to exhibit substantial visual variation, while class imbalance under-represents clinically important high-grade patterns. The effect of class imbalance is not limited to training data: it can re-emerge during unsupervised test-time adaptation, further suppressing rare but diagnostically important classes even as overall performance improves. To address these coupled deployment challenges, we propose Anti-Adversarial Consistency with Test-Time Adaptation (AAC-TTA), an unsupervised test-time adaptation framework for imbalanced histopathology classification under scanner-induced domain shift. AAC-TTA probes local prediction stability using small entropy-guided adversarial and anti-adversarial perturbation views, and derives a more stable target-domain training signal from the views that pass a reliability check. It further combines agreement-based reliability gating with constrained adaptation of lightweight model components, so that adaptation is not driven by unreliable pseudo-labels. Experiments on AGGC22-derived H&E-stained prostate histopathology patches for automated Gleason grading, comprising 95,400 training and 47,700 test patches across three scanner domains (Akoya, Philips, and KFBio) show that AAC-TTA achieves the highest macro-F1 among representative test-time adaptation baselines while substantially improving recognition of the under-represented high-grade class. Rather than artificially constructing separate no-, moderate-, and high-imbalance scenarios, we evaluate the single G5-minority imbalance that arises naturally in the AGGC22 data, since G5 is both intrinsically rare and the most diagnostically critical high-grade pattern. Our results indicate that medical test-time adaptation methods should be evaluated not only by overall performance but also by their ability to preserve minority-class recognition under domain shift.