Advancing brain tumor segmentation via Vision Mamba and soft labels

Abstract

Brain tumors are a result of the abnormal growth of cells in the brain. Gliomas are the most common type of brain tumors in adults and are particularly aggressive, as they impair brain functions and often lead to death. The accurate and fast diagnosis and monitoring of brain tumors can significantly improve a patient’s chances of survival. Fortunately, magnetic resonance imaging (MRI) provides a non-invasive means to assess brain tumors and their progression both qualitatively and quantitatively. However, manual analysis of MRI scans is time-consuming and labor-intensive. To address this challenge, machine learning techniques can be used to automatically segment brain tumors in magnetic resonance (MR) images, thereby significantly improving diagnostic efficiency. The task of brain tumor segmentation is nonetheless challenging, due to the large size of these scans and ambiguities stemming from tissue contrast and image quality. Both of these issues pose significant obstacles to the advancement of artificial intelligence methods for this task. This thesis aims to address these two challenges by developing a lightweight and robust methodology for brain tumor segmentation.

The proposed solution consists of an efficient neural architecture based on Vision Mamba, whose computational complexity increases linearly with the input size, and a label smoothing method that considers both positional and visual information in MR images to quantify the lack of confidence in the ground truth. The thorough experiments on two public datasets, BraTS 2020 and LGG FLAIR, demonstrated the advantages of the presented methodology over commonly employed approaches in this domain and highlighted two key findings. Firstly, integrating suitable token extractors enhanced the performance of Vision Mamba on brain tumor segmentation while maintaining its efficiency. Secondly, the introduced label smoothing technique improved the quality of segmenting brain tumors across various deep learning architectures. As a result, the developed solution achieved competitive performance in brain tumor segmentation and showed promising potential for clinical applications.