DARE: Domain-Aware Representation Enhancement for data-efficient histopathological tissue classification

Deep learning (DL) has achieved remarkable progress in medical image analysis. However, its success critically depends on large-scale, expert-annotated datasets, which are costly and impractical to obtain in clinical practice. This limitation is further aggravated by the inherent heterogeneity of medical imaging across institutions, making it even more challenging to curate reliable and consistent datasets. Therefore, DL models trained on a single dataset often exhibit limited generalizability and poor transferability especially under the limited training data scenario. To address these challenges in the Histopathological Tissue classification task, we propose a novel framework named Domain-Augmentation via Representative Enrichment (DARE), which integrates two data augmentation strategies within an advanced lightweight deep broad learning system for targeted domain adaptation. First, we analyze the source-domain dataset in the feature space and argue that, a single global statistic is insufficient to fully characterize the source-domain distribution under limited training data scenario. Therefore, we introduce DARE†, a pseudo-domain partition-based feature-space augmentation strategy. By partitioning the limited source-domain data into multiple pseudo-domains via the thought of sampling, DARE† extracts diverse statistical representations that enrich domain variability and thereby improve generalization to the target domain. Second, recognizing that the original image space fundamentally shapes the feature space, we further design DARE*, an image-level augmentation scheme. DARE* directly increases the diversity of source-domain images, complementing the feature-space augmentation and further enhancing the model’s robustness to domain shift. Extensive experiments demonstrate that our proposed methods achieve consistent performance improvements on target domains under extremely limited source-domain training scenarios. The source code is available at: https://github.com/AndyRong921/DARE