Attention-Enhanced Hybrid U-Net With Multi-Modal Mri Feature Fusion for Brain Tumor Segmentation

Reliable outlining of brain tumours on magnetic resonance imaging (MRI) is a demanding research problem, because the appearance of tumour subregions varies between imaging sequences and from one patient to the next.This paper reports an attention-enhanced hybrid U-Net for multi-modal brain-tumour segmentation.The network takes four co-registered MRI sequences as input (T1, contrast-enhanced T1 (T1CE), T2 and fluidattenuated inversion recovery (FLAIR)) and assigns every pixel to one of four classes: background, necrotic/non-enhancing tumour core, oedema, or enhancing tumour.Its design joins a convolutional encoderdecoder, skip connections filtered by attention gates, and a Transformer-based bottleneck.We train and test on the BraTS 2021 training collection, divided at patient level into 1,000 training, 125 validation and 126 test cases.Training ran for five epochs on 128 × 128 axial slices with the AdamW optimiser and a loss that adds cross-entropy to soft Dice.In a preliminary held-out test that draws one deterministic tumour-positive axial slice from each test patient, the pooled macro Dice reached 0.7570 and the macro IoU 0.6157.The class-wise Dice scores were 0.6422 for the necrotic/non-enhancing core, 0.8042 for oedema and 0.8246 for enhancing tumour.These numbers are pooled over sampled slices; they are neither full-volume nor patient-averaged BraTS benchmark scores.We document the architecture, the data protocol, the observed performance and the limitations so that later experiments can be reproduced.