PVRSAP-Net: Probabilistic Voxel Refinement and Spatial Attention Pyramid Learning for Three-Dimensional Multiregion Brain Tumor Segmentation

Abstract

Accurate three-dimensional delineation of glioma subregions from multimodal magnetic resonance imaging remains difficult when tumor margins are weak, lesion volumes are small, acquisition intensity varies across subjects, or a coarse segmentation prior contains spatially isolated errors. The supplied research synopsis proposes a voxel-based framework that joins probabilistic threshold refinement, adaptive multiscale descriptors, spatial-attention pyramid learning, residual convolution, and multiregion classification. This paper converts that concept into a single testable method named PVRSAP-Net. Its first mechanism is bounded edge-aware probabilistic voxel refinement, which updates uncertain tumor probabilities with neighborhood evidence, image-derived conductance, projection to the valid probability interval and nested-region constraint set, and explicit relations among enhancing tumor, tumor core, and whole tumor. The refined probabilities join normalized multimodal intensity, local texture, and contextual descriptors. A windowed spatial-attention pyramid then models nonlocal dependence at several resolutions, and a residual three-dimensional decoder predicts nested tumor regions with confidence estimates. Training uses a region-weighted Dice, cross-entropy, boundary, and calibration objective. Three algorithms, three conditional mathematical propositions, computational analysis, safeguards, and a leakage-resistant BraTS evaluation protocol are specified. Since raw clinical volumes and executable baseline predictions were not supplied, no clinical superiority claim is made. A seeded controlled simulation of 20 synthetic four-channel multimodal volumes, each represented on a 483 voxel grid, tests only the internal refinement behavior. In that simulation, PVRSAP-Net refinement attained pooled Dice 0.9091, IoU 0.8382, sensitivity 0.9289, HD95 1.2316 voxels, Brier score 0.0040, and expected calibration error 0.0142. The corresponding raw probabilistic prior attained Dice 0.8551 and HD95 7.1811 voxels. Removal of edge conductance reduced Dice to 0.8844; removal of anatomical hierarchy projection reduced Dice to 0.8802. These results support numerical and structural behavior under declared synthetic assumptions, not patient-level efficacy.