Automated symmetry analysis in neonatal brain MRI remains a critical challenge for early detection of developmental abnormalities, yet existing deep learning approaches struggle with the unique characteristics of neona- tal imaging. This paper presents a systematic evaluation of Siamese neural network architectures for this task, comparing five state-of-the-art back- bone networks: ResNet, ResNeXt, MobileNet, VGG, and EfficientNet. We develop a novel training methodology utilizing controlled asymmetry sim- ulation across 3,150 annotated axial brain views from a novel dataset. Our approach implements a progressive training protocol with asymmetries ranging from 1 mm2 to 20 mm2, enabling an evaluation of each architec- ture’s detection capabilities. Comprehensive experiments demonstrate that VGG-based architectures achieve superior performance, with detection accuracy ranging from 76% for 7mm2 asymmetries to 99% for asymme- tries above 13.5 mm2. Notably, our analysis reveals that models trained on medium-range asymmetries demonstrate better generalization capabilities across different scales. By providing a valuable resource for symmetry analysis in neonatal neuroimaging and providing quantitative insights into architectural design choices for medical image deep learning, this work contributes to the development of more sensitive and robust diagnostic tools in neonatal diagnosis and care.