Abstract / Summary
Introduction: Abdominal aortic aneurysms (AAAs) represent a rare but critical surgical emergency, as ruptures carry significantly high mortality rates, making early detection essential for prevention. Machine learning models offer promising capabilities for aneurysm detection; however, their performance is constrained by the scarcity of annotated imaging datasets. Materials and methods: This comparative study employed an intentionally simple experimental framework to isolate the effect of synthetic data augmentation on classification performance. A deep convolutional generative adversarial network (GAN) was implemented to generate synthetic images and expand a limited dataset of 163 computed tomography (CT)-derived three-dimensional abdominal aorta models. Synthetic aneurysm samples were combined with original images and expanded through augmentation to construct progressively larger datasets comprising 250, 500, 1000, 2500, 5000, 7500, and 10,000 images. A Residual Network 50 (ResNet50) convolutional neural network was trained and evaluated at each dataset size. Gradient-weighted class activation mapping (Grad-CAM) heatmaps were generated to assess model interpretability and identify image regions contributing to classification decisions. Results: The ResNet50 model achieved classification accuracies ranging from 97% to 99% across all dataset sizes, with correspondingly high precision, recall, and F1 scores for both aneurysmal and normal classifications. Grad-CAM visualization produced mixed results regarding anatomical localization of clinically relevant features. Conclusions: Synthetic dataset expansion through GANs represents a viable approach for developing high-accuracy classification models while overcoming limited medical imaging data availability. However, the mixed interpretability results warrant further investigation to ensure clinical applicability and physician trust in model predictions.