Abstract
This study conducts a comprehensive analysis of various configurations of the UNet + ConvNeXt Tiny architecture for breast tumor segmentation. We assess the influence of data augmentation, skip connections, attention mechanisms, and dropout rates on segmentation performance utilizing the BUSI dataset, subsequently conducting cross-dataset evaluation on the BUS-UCLM dataset. Our findings indicate that the integration of attention mechanisms markedly improves metrics including F1 score, precision, IoU, recall, and Dice coefficient. Data augmentation alone did not consistently enhance model performance and occasionally caused instability, however its integration with attention processes frequently resulted in significant improvements. Dropout was helpful in mitigating overfitting, with a rate of 0.5 consistently yielding enhanced generalization in both in-domain and cross-domain assessments, especially when combined with strategically positioned attention modules. Models using attention mechanisms and a dropout rate of 0.5 attained the optimal equilibrium between accuracy and recall, emphasizing the necessity of balancing regularization intensity with architectural intricacy. Nevertheless, our results underscore the potential for certain augmentation techniques-particularly those that employ aggressive geometric transformations-to distort anatomical structures in ultrasound images, resulting in a decline in the quality of segmentation. This highlights the necessity for meticulous, domain-specific selection of augmentation strategies in medical imaging. This study emphasizes the necessity for meticulous selection of architectural elements and regularization techniques to enhance medical image segmentation and offers direction for future investigations in resilient deep learning applications for clinical imaging.