Umeå University's logo

Generalization is a central goal in machine learning. Good performance on development data does not always mean reliable performance in practice. A central clinical motivation for this thesis is to improve model performance for pediatric patients, whose data are often underrepresented relative to adult patients. This reflects a broader challenge in machine learning: reliable generalization depends on both methodological choices and how well different groups are represented in the training data. Against this background, the thesis asks what methodological pitfalls affect observed performance and how they are addressed in current machine learning research, how existing latent-space oversampling methods address imbalance, how the choice of learned representation affects their effectiveness, and how learned representations can guide generative models for minority-sample generation.

To address the first question, the thesis identifies 14 common methodological pitfalls across the machine learning pipeline, examines their effects on observed model performance through controlled experiments, and studies how clearly they are reported in published work. The thesis then addresses the remaining questions by focusing on data imbalance. Its main approach is synthetic oversampling in learned representation spaces, where new minority representations are constructed and then used to generate synthetic samples. It first studies DeepSMOTE, an existing latent-space oversampling method based on an autoencoder, and analyzes its penalty term and its effect on the learned latent space. It then applies latent-space oversampling to adult-pediatric imbalance in brain tumor segmentation using a cohort-conditioned StyleGAN2 model to generate synthetic pediatric MRI and segmentation-mask pairs. Finally, it studies representation-conditioned diffusion models for class-imbalanced image classification and compares different strategies for constructing minority representations.

The results show that methodological choices throughout the machine learning pipeline can notably affect observed model performance, and that such issues are not always reported clearly in published studies. For latent-space oversampling, the results show that balancing the classes is not enough by itself. What matters is also how the additional synthetic minority samples are created and on the representation space in which this is done. In the adult-pediatric setting, latent-space oversampling improved segmentation performance for the underrepresented pediatric cohort while maintaining performance for adult patients. Overall, synthetic data can improve generalization for minority classes and subgroups. However, the benefit depends on the representation, the generation strategy, and the used training procedure.