Architecture Matters for Small Tumors: Compute-Matched Comparison of U-Net, Attention U-Net, and UNet++ for Brain Tumor Segmentation

Introduction: Brain tumor segmentation from magnetic resonance imaging (MRI) is an important problem in computer-aided diagnosis, treatment planning, and image-guided surgical intervention. Accurate segmentation remains challenging because tumors often exhibit highly heterogeneous morphology, unclear boundaries, low contrast relative to surrounding tissues, and severe imbalance between tumor and non-tumor pixels.

Methods: In this work, we developed and systematically optimized a deep learning-based Attention U-Net segmentation framework using the BraTS 2020 benchmark dataset. The training pipeline incorporates a hybrid Binary Cross-Entropy and Dice loss function, class-aware data augmentation strategies, test-time augmentation, and extensive hyperparameter optimization. Using the best optimized configuration, we performed a controlled compute-matched comparison of U-Net, Attention U-Net, and UNet++ under an identical training pipeline, fixed computational budget, and tumor-size-stratified evaluation setting.

Results: The optimized Attention U-Net achieved an overall Dice score of 0.9074 and a tumor-only Dice score of 0.8081. Although all three architectures achieved comparable aggregate segmentation performance, UNet++ demonstrated significantly stronger robustness on clinically important small-tumor cases, improving small-tumor Dice performance from 0.330 to 0.400 while reducing HD95 boundary error from 55.6 to 46.8 pixels compared with Attention U-Net.

Conclusions: Our findings show that aggregate evaluation metrics can hide clinically meaningful architectural differences, and that dense nested skip connections improve challenging small-lesion segmentation performance under realistic computational constraints for medical imaging applications.