| Abstrakt: |
Postoperative brain magnetic resonance images (MRI) is pivotal for evaluating tumor resection and monitoring post-surgical changes. The segmentation of surgical cavities in these images poses challenges due to artifacts, tissue reorganization, and heterogeneous appearances. This study explores challenges and advancements in postoperative brain MRI segmentation, examining publicly accessible datasets and the efficacy of various deep learning models. The analysis focuses on different U-Net models (U-Net, V-Net, ResU-Net, attention U-Net, dense U-Net, and dilated U-Net) using the EPISURG dataset. The training dice scores are as follows: U-Net 0.8150, attention U-Net 0.8534, V-Net 0.7602, ResU-Net 0.7945, dense U-Net 0.83, dilated U-Net 0.80. The study thoroughly assesses existing postoperative cavity segmentation models and proposes a finetuning approach to enhance the performance further, particularly for the best-performing model, attention U-Net. This fine-tuning involves introducing dilated convolutions and residual connections to the existing attention U-Net model, resulting in improved results. These improvements underscore the necessity for ongoing research to select and adapt efficient models, retrain specific layers with a comprehensive collection of postoperative images, and fine-tune model parameters to enhance feature extraction during the encoding phase. [ABSTRACT FROM AUTHOR] |
