Brain tumor segmentation using dual-stream multiscale 3D-UNET with dense net and spatial attention.

This research work proposes an improved approach for brain tumor segmentation using a Dual-stream multi-scale 3D-UNET network integrated with DenseNet and spatial attention mechanisms. Identifying brain tumors in multichannel Magnetic Resonance Imaging (MRI) is crucial for the diagnosis and subsequent treatment. Some of these challenges include the unpredictable nature of the tumor's appearance and the need to define the subregions of the tumor accurately. Since our objectives involve updating the segmentation models for various tumor classes to better align with the BRATS 2020 MRI standard dataset, which contains MRI scans labeled by experts, we want to focus on the seven categories, including edema, enhancing tumors, and core tumors. The proposed model employs a dual-stream approach to recognize multiple input formats simultaneously, utilizing DenseNet to extract features and apply spatial attention, thereby focusing on areas of interest within the imaging data. The model achieved remarkable Dice scores, outperforming baseline results: 99.12 for improving tumors, 99.99 for swelling, and 99.88 for core tumors. The typical Dice score was 99.99 for all models on the BRATS 2020 dataset. The results are further validated when compared with the more recent BRATS 2021 set, where the accuracies are 99.45%, 98.24%, and 99.55% for the same categories, respectively. Future efforts will focus on how to adapt the suggested model to real-time data, how to consider other forms of generalization of unsupervised learning, and how clinician feedback can be incorporated to improve the interpretability and applicability of the approach to a clinical setting.