The impact of loss functions and optimizers on U-Net for aneurysmal ascending aorta segmentation

Aneurysmal ascending aorta (AsAA) is a life-threatening condition that requires precise segmentation for diagnosis and surgical planning. Manual segmentation of AsAA using Computer Tomography (CT) imaging is time-consuming and labour-intensive, making it unsuitable for routine clinical practice. This study explores the application of deep learning, specifically U-Net architecture with an attention gate, for automated segmentation of AsAA from CT angiography images. The performance of various loss functions and optimizers is examined, including a custom loss function that combines Binary Cross-Entropy (BCE) and weighted Dice loss and the Dice Focal loss, to address the class imbalance issue. Additionally, the interaction between two commonly used optimizers, Adaptive Moment Estimation (Adam) and Stochastic Gradient Descent (SGD), is investigated. Results show that compound loss functions of BCE and Dice, paired with the SGD optimizer, outperformed other configurations, yielding a median Dice score of 0.98 and Hausdorff distance around 2.5. The SGD optimizer showed more stable performance compared to Adam during the validation process and higher segmentation performance in the test set. These findings underline the critical role of optimizer and loss function selection into the design of U-Net-based models for improving AsAA segmentation accuracy. Future research should explore enhancements to the model architecture and validation on larger, more diverse datasets for broader clinical application.