Brain Tumor Segmentation using CT Hybrid

Abstract

One of the most recent deep learning innovations is the use of hybrids of transformers-CNN architectures for medical image segmentation. Given that computer vision and medical imaging share certain similarities, it is only logical to wonder whether the transformer hybrids can change the way that medical imaging is done. Brain tumor segmentation, especially one of its subregions i.e., Enhancing Tumor, is quite challenging to segment. In his paper, we have used a modified version of the SWIN UNETR model, CT Hybrid. We have done transfer learning and freezing some of the model’ s layers while tuning some hyperparameters, we have achieved SOTA performance within a few epochs, utilizing minimal time and computational resources. We retrained the CT model using the BraTS 2023 dataset. For enhancing tumor our score showed improvement and outperformed recent best performing UNet and transformer architectures.