Fused Hybrid Training Samples Through Synthetic Anomaly Generation for Optimized Model Training

Deep neural networks for classification and segmentation of complex images require a sufficient amount of annotated data. Model performance depends significantly on the number of samples and the class balance within the dataset. In particular, anomalies and edge cases are often underrepresented, as they occur less frequently than control samples under normal conditions. However, collecting and annotating such data is usually associated with significant financial and time costs, as seen in applications like quality control in production. Moreover, in the medical sector, additional challenges such as data privacy can hinder the compilation of sufficient datasets. Recent studies show that generated synthetic samples, in combination with real samples, have the potential to expand underrepresented classes, increase variance and thereby improve model performance. However, generating realistic and complex multidimensional data is challenging and typically requires extensive domain knowledge.This work introduces a simple, automated concept to expanding complex image datasets with generated and fused hybrid training samples, providing an intermediate solution between purely real and entirely synthetic samples. Using a 3D convolutional variational autoencoder, new segmented variations of anomalies were generated and fused with real control samples. A comparison study was conducted, with the expanded hybrid dataset, along with the isolated real data, to examine the impact of generated hybrid samples. Our results demonstrate that the inclusion of synthetic anomalies enhances the training process and improves detection performance on 3D medical MRI data. The most promising models, trained on single sequences, achieved consistent or improved results across all evaluation metrics compared to models trained exclusively on real data. Since the proposed concept shows promising results, hybrid sample generation could become a valuable tool for data preparation and augmentation to optimize complex datasets and improve model training.