Dual-Stage IoU Optimized Multiple Scale Attention Networks for Brain Tumour Segmentation Using Cross Modal and Structural Learning Process

Abstract

For clinical diagnosis, treatment planning, and prognosis, the accurate segmentation of brain tumours from MRI images is an area in need of utmost research sets. On the downside, with the deep learning-based segmentation models, key problems of limited generalizability, low boundary precision, and poor exploitation of the multi-modal MRI data samples have arisen in process. Most of the present models do not seem to suitably characterize cross-modal contextual dependencies, are hesitant when it comes to fine-grained boundary segmentation, and often depend on static loss functions, incapable of adapting to rapid intra-class and inter-class variabilities throughout the training process. To contend with these challenges, we put forth a new Multi-Scale Attention-Driven CNN Framework for Brain Tumour Segmentation with Dual-Stage IOU Optimization, which encompasses five analytically motivated and technically new modules. First, the CMSCA mechanism dynamically integrates features across MRI modalities by evaluating inter-modal consistencies at multiple spatial resolution levels, thus enhancing cross-modal integration. Second, the Dual-Domain Gradient Agreement Loss (DDGAL) holds edge-aware supervision in the spatial domain and frequency domain to increase structural accuracy. Third, the Temporal Feature Accumulation with IOU Memory Units (TFA IMU) tracks learning dynamics across training epochs to enable adaptive feature re-weighting according to historical IOU patterns. Fourth, Graph-Enforced Regional Attention with Structural Consistency Loss (GERA-SCL) embeds superpixel-based structural priors via graph convolutions to assure label consistency and topological coherence within the tumour region. Finally, based on adaptive pseudo labelling and IOU-driven confidence estimation, the Adaptive Dual-Stage Pseudo IOU Feedback Refinement (ADP IFR) is proposed to enhance the boundary accuracy on unlabelled data samples. The suggested framework demonstrates sustained improvements in segmentation accuracy, boundary finesse, and generalization ability sets. The Dice score evidence shows an increase of +4.2%, an increase of +4.1% in IOU, and a decrease of −6.5 px in Hausdorff Distance, implying marked superiority over the contemporary state-of-the-art methods within the realms of labelling and semi-supervised learning in the process.