Publications

Medical Image Segmentation System, Bachelor's project, 2023 [pdf] [link]
Supervisor: Maryam A. Mazlaghani
abstract
Rapid advances in the field of medical imaging are revolutionizing medicine. For example, the diseases diagnosis with the help of computers, where the segmentation of medical images plays an important role, has become more accurate. Although CNN-based methods have achieved excellent performance in recent years, but due to the intrinsic locality of convolution operations, they cannot learn explicit global and long-range semantic information well. Given the increased interest in self-attention mechanisms in computer vision and their ability to overcome this problem, the TransUNet architecture was proposed as the first medical image segmentation framework using Vision Transformer as a strong encoder in a U-shaped architecture. TransUNet achieves good results compared to different architectures; therefore, in this project, we use it as the base model that has a hybrid CNN-Transformer architecture. this architecture is able to leverage both detailed high-resolution spatial information from CNN features and the global context encoded by Transformers. All experiments are conducted on Kvasir-SEG, CVC-ClinicDB and Ph2 datasets. First, we reproduce the results in the original paper, and then we proceed to improve the architecture by making appropriate changes and check the results. Some of these changes have been successful and others have been unsuccessful. Finally, we created a web-based system based on the new architecture.