NeuroFocusNet: An Attention-Enhanced 3-D U-Net for Multimodal MRI Brain Tumor Segmentation

Abstract

Manual delineation is time-consuming, and inter-reader variability is high, making accurate delineation of glioma subregions in multimodal magnetic resonance imaging (MRI) important for treatment planning and longitudinal assessment. Current automatic techniques have limitations in identifying small enhancing regions, maintaining irregular tumor boundaries, and maintaining performance in the event of changes in image quality or input availability. This study proposes a 3D U-Net-based segmentation framework, NeuroFocusNet, which synchronizes spatial attention, channel recalibration, and attention-gated skip connections within a single volumetric encoder-decoder model. The framework also features intensity preprocessing using standardization, tumor-biased patch sampling, and a composite loss function comprising Dice, cross-entropy, focal, and boundary-aware loss terms. On the BraTS 2023 dataset, NeuroFocusNet achieved Dice scores across the whole tumor (WT), tumor core (TC), and enhancing tumor (ET) regions of 0.9268, 0.9081, and 0.8707, respectively, with corresponding mean Hausdorff-95 distances of 1.68 mm, 1.36 mm, and 2.10 mm, respectively. A performance reduction was demonstrated when adding simulated Gaussian noise to the input, and was further demonstrated when a modality mask was applied during inference. Furthermore, 27 randomly selected cases were submitted to 3 different neuroradiologists for a preliminary evaluation of anatomical plausibility and correction requirements. This reader study is not meant to be a prospective or multicentre clinical validation of the framework. The results are encouraging when considering the feasibility of using NeuroFocusNet as a technically promising segmentation framework, and the need for externally controlled benchmarking, multicentre assessment, and prospective clinical assessment.