An Optimal U-Net++ Segmentation Method for Dataset BUSI
Jiaqi Shang, Yinyi Lai · 2024
The ability to correctly diagnose cancer is crucial for determining prognosis and planning appropriate treatment. At present, there is still a lack of research on various backbone structures within U-Net. This study analyzes the various backbone architectures of U-Net++. By carefully selecting the basic model, the impact of various feature extraction encoders on the results was explored. In order to improve computational efficiency and accuracy, optimization of input size and batch size has been introduced. In addition, by combining transfer learning with pre-trained models, a strategy to improve model performance was proposed: by changing the base model and transfer learning, the base model was ResNet18, the input size was (96,96), and the batch processing was 48. The best U-Net++ model for segmentation on the BUSI dataset was identified using transfer learning as the ImageNet. The structured research methods and conclusions proposed in this study significantly promote the segmentation of breast cancer and provide a promising direction for future research and clinical application.