Automated Brain Tumor Segmentation in Multimodal MRI Scans Using a Multi-Encoder U-Net Model

Gliomas are recognized as the prevailing type of primary brain tumors. The automated segmentation of these tumors from 3D magnetic resonance imaging (MRI) plays a key role in the diagnosis, treatment planning and the monitoring of this disease. This paper presents a novel 2D U-Net architecture designed to automatically detect and segment three different tumor sub-regions, (edema, enhancing tumor, and necrosis) in preoperative multimodal 3D brain MRI scans. The architecture of our convolutional neural network (CNN) model consists of three encoders, facilitating feature extraction from four different modalities of brain MRI images at multiple scales, as well as a decoder to efficiently upscale learned features while preserving spatial information. The proposed CNN model was experimented and assessed on 5880 multimodal brain MRIs from BraTS'2023 training and validation datasets, representing 1470 different adult subjects having high and low-grade gliomas. It achieved an average Dice score of 0.92 for whole tumor (WT), 0.92 for tumor core (TC), and 0.90 for enhancing tumor (ET), when evaluated on the BraTS'2023 training dataset. On the validation dataset, the model's mean Dice scores for WT, TC, and ET were 0.89, 0.88, and 0.73, respectively.