An Efficient Multilevel-Fusion 3-D Framework for Brain Tumor Segmentation in Multimodal MRI

Abstract

Accurate volumetric delineation of glioma subregions from multimodal magnetic resonance imaging (MRI) supports quantitative assessment, treatment planning, and longitudinal follow-up. This paper presents a reproducibility-aware rewrite of a multilevel-fusion framework that integrates three established ideas: robust modality-wise intensity conditioning, a 3-D residual inception encoder-decoder with guided attention and multiscale fusion, and fuzzy graph-cut refinement of voxelwise posterior probabilities. The formulation preserves local detail through residual paths, models long-range context at the bottleneck, and imposes spatial coherence during refinement. Three algorithms are specified in standard pseudocode, followed by mathematical results for range preservation, convex-gated fusion, probability normalization, descent of the alternating refinement energy, and the limits of those guarantees. The experimental record provided with the source manuscript identifies BraTS 2020, 128 x 128 x 128 inputs, Adam optimization, 100 epochs, and an NVIDIA A100. It reports aggregate Dice, intersection-over-union, sensitivity, specificity, accuracy, recall, precision, and HD95 values. These values are reproduced as reported and are not presented as independently re-estimated results, since subject-level predictions, split identifiers, repeated runs, and code were not supplied. The paper therefore distinguishes architectural analysis from empirical evidence, documents reporting inconsistencies, and defines the minimum validation package needed for a defensible comparison.