Automated Detection and Segmentation of Glioblastoma in MRI Using Multi-Level Diffusion Transformer U-Net

Detecting malignant tumors at an early stage is crucial for effective treatment and an improved prognosis of brain cancer. Radiologists use MRI imaging modality to detect and diagnose glioblastoma. For efficient detection, numerous deeplearning methods have been proposed in the past decade. Detecting brain tumors in MR images is challenging owing to the heterogeneity in anatomical structures across patients, and difficulties posed by low contrast, inhomogeneous intensity levels, and artifacts. In this paper, we proposed a multi-level Diffusion Transformer U-Net (mDT-UNet) with a modified cross-attention module. Our model extracts features from both global and local levels of the MR images and manual segmentations providing a wider range of contextual information. Our proposed model attained an average accuracy of 95% on the BraTS 2020 dataset whereas U-Net achieved 81% and Swin Transformer achieved 86% accuracy. In addition, we evaluated our model on various benchmark datasets; CPM-RadPath, UCSF PDGM, EGD, and IvyGAP. With appropriate pre-training on large medical datasets, our model can be optimized for various medical image segmentation tasks.