School of Engineering Department of Computer Science and Engineering 106 Artificial Intelligence Against COVID-19 Supervisor: CHEN, Hao / CSE Student: OW YONG, Chee Seng / CPEG Course: UROP2100, Spring The segmentation of brain tumors is an important task in medical image analysis and can help in the diagnosis and treatment of patients. In this research, I explore the use of a 3D U-Net model for semantic segmentation of brain tumors on the BraTS 2020 dataset. The proposed method involves preprocessing the dataset by scaling the images using a MinMax scaler, combining them into a multi-channel volume, cropping images to remove blank regions and dropping images that have less than 1% annotation. Our proposed model achieved an accuracy of 0.991 and intersection over union (IOU) of 0.805 on the training set and an accuracy of 0.981 and IOU of 0.652 on the validation set. We also explored the use of attention U-Net and attention residual U-Net as potential improvements to our model. Our results show that U-Net is an effective model for brain tumor segmentation with potential for further improvement using attention mechanisms. Artificial Intelligence Against COVID-19 Supervisor: CHEN, Hao / CSE Student: ZENG, Lingqi / DSCT Course: UROP1100, Spring This report presents a study and experimentation of a memory bank-based estimation scheme for test-time adaptation in segmentation tasks. The proposed method involves adapting the test-time learning rate based on the estimated discrepancy for each test sample, which has shown an improved model generalization across different domains. Results show a significant improvement in the average Dice score after adaptation, leading to increased performance and robustness in segmentation tasks.
RkJQdWJsaXNoZXIy NDk5Njg=