Accurate brain tumor segmentation on BraTS 2021 dataset based on 3D U-Net: a lightweight clinical auxiliary diagnosis system

Abstract

In the diagnosis and treatment of intracranial tumors, magnetic resonance imaging (MRI) is widely regarded as the gold standard for tumor detection and classification due to its non-invasiveness, which is crucial for surgical planning and prognosis evaluation. However, the current clinical diagnosis of intracranial tumors mainly relies on manual identification and delineation of MRI images by radiologists, resulting in a huge workload. Diagnostic results are greatly affected by physicians’ subjective experience and visual sensitivity, which may interfere with surgical planning and introduce uncontrollable risks. 2D UNet loses 3D spatial information, suffers from long training time and weak contextual correlation, and thus cannot meet the precise segmentation and prediction requirements of medical images. 3D V-Net can effectively process 3D image features and further enhance image processing capabilities, but it has problems such as complex structure, slow convergence, high hardware requirements, and insufficient boundary fitting accuracy, making it difficult to meet the lightweight deployment demands in clinical practice. Therefore, this study selects the 3D U-Net model with skip connections and builds a lightweight clinical auxiliary diagnosis system to realize front-end and back-end deployment.