An Automatic Brain Tumor Segmentation Using 3D LSRA SegFormer

Abstract

Cancer cells grow abnormally, which should not occur in the human body. In the case of brain tissue, this abnormal growth leads to the formation of a mass known as a brain tumor. The identification of tumor regions in clinical diagnosis has been performed using established methodologies and scientific protocols that primarily rely on the expertise of experienced clinicians. Competent medical and surgical strategies enhance survival and treatment options for individuals with brain tumors. Unfortunately, achieving precise segmentation is challenging due to the variety of shapes and appearances of brain tumors. Consequently, the variation in tumor segmentation between doctors can lead to bias and variability in diagnoses, potentially resulting in different conclusions for the same tumor among pathologists. Hence, there is a need for automated segmentation techniques to expedite the process while also improving diagnostic efficiency, accuracy, and consistency. Deep learning has gained popularity and is renowned for its effectiveness in segmenting brain tumors. This paper proposes an efficient and lightweight transformer-based 3D MRI brain tumor segmentation method called "3D LSRA SegFormer." The feature extractor employs a Pyramid Vision Transformer structure with linear spatial reduction attention (LSRA) and connects to a lightweight multi-layer perceptron decoder at various stages through skip-connections. The feature extractor layer learns to interact with global information in 3D MRI images, while the decoder works to upsample the feature maps to generate segmentation. LSRA aims to reduce the computation of large feature maps, achieve better accuracy, and produce an efficient model. We validate the successful results of the 3D LSRA SegFormer method for the task of multi-modal 3D brain tumor segmentation in the Multi-modal Brain Tumor Segmentation Challenge (BraTS) 2021 dataset.