SSGNet++: Stage-Aware Generative Augmentation and Ensemble-Guided Pseudo-Labelling for Medical Imaging

Limited annotation and class imbalance remain important challenges in medical image analysis. SSGNet++ examines class-specific StyleGAN3 augmentation for classification and an iterative pseudo-labelling pipeline for segmentation, combining predictions from heterogeneous multi-base and multi-head models with optional mask refinement. We evaluate four classification tasks and four segmentation tasks drawn from six public datasets. In the classification experiments, the best tested synthetic-data settings increased accuracy by 3–7 percentage points over real-only training in several dataset settings, although the preferred augmentation ratio varied and larger ratios could reduce performance. The initial segmentation analyses examined single models, ensembles, iterative pseudo-label training, and refinement stages, but reused validation partitions for both model selection and performance reporting; these results are, therefore, exploratory and may be optimistic. Four-seed follow-ups comprised a post-hoc internal split (112 runs) and a partial external-cohort evaluation (64 runs); multi-head gains were uncertain and refinement effects varied by dataset. These follow-ups remain limited by the post-hoc internal split and incomplete external comparisons.