Comprehensive Evaluation of L2 Regularization in UNet-ResNet50 and UNet-VGG16 Models with Diverse Image Transformations

Automated medical image segmentation is essential for clinical diagnosis and reduces the reliance on manual annotation. Deep learning models exhibit high accuracy but suffer from overfitting and sensitivity to image perturbations. Regularization methods, including L2 weight decomposition, mitigate these issues by limiting model complexity. This study evaluates the impact of L2 regularization on segmentation robustness using the UNet, UNet-ResNet50, and UNet-VGG16 architectures. The effect of regularization on model generalization is evaluated under structured perturbations. Chest X-ray datasets from Montgomery and Shenzhen were used. Segmentation models were trained with and without L2 regularization at λ = 10^-4 and λ = 10^-3. Gaussian noise and blur were applied for testing. The novelty of the work lies in the analysis of the correlation between the L2 regularization parameters and the robustness of the models to noise and blurring of the input data, revealing optimal configurations for preserving segmentation accuracy under conditions of image degradation. In contrast to previous studies that focus on improving segmentation accuracy, this paper conducts an extensive study of how L2 regularization affects the resilience of different architectures to distortion in structured input signals. Regularisation at the level of λ = 10^-4 enhanced reliability without substantially affecting the segmentation accuracy of the original images. However, this moderate regularisation resulted in a minor compromise, the Dice estimate for the UNet-VGG16 model decreased by 1.6 percentage points on these pristine images. The regularised models demonstrated a tendency to exhibit reduced degradation when subjected to the effects of noise and blurring. It is found that the UNet-VGG16 architecture with a regularization factor of λ=10⁻⁴ achieves an optimal balance between a Dice segmentation accuracy of 0.94 and robustness to image perturbations. These results contribute to the methodology of optimizing deep neural networks for medical segmentation tasks.