FreqTextSeg3D: Frequency-Aware Text-Guided 3D Brain Tumor Segmentation via Hierarchical Semantic Decomposition and Dual-Path Cross-Modal Fusion

Abstract

Accurate 3D brain tumor segmentation is critical for diagnosis, treatment planning, and prognosis monitoring. Although recent vision-language methods have begun to incorporate radiology reports into segmentation pipelines, they typically treat clinical text as a monolithic embedding, ignoring its rich sub-semantic structure. We present FreqTextSeg3D, a novel framework that introduces three tightly coupled contributions: (i) Hierarchical Semantic Text Decomposition (HSTD), which separates a radiology report into three specialized streams (anatomical, morphological, and diagnostic) via learned soft gates with a diversity loss; (ii) Frequency-Aware Cross-Modal Fusion (FACMF), a dual-path module that combines spatial cross-attention with 3D Fourier transform (FT) amplitude/phase modulation to inject structured text priors into visual features; and (iii) Region-Text Contrastive Alignment (RTCA), a symmetric InfoNCE loss that anchors each text stream to its corresponding tumor sub-region. Evaluated on BraTS2020 under 5-fold cross-validation, FreqTextSeg3D achieves mean Dice scores of 91.36% (Whole Tumor), 86.49% (Tumor Core), and 77.73% (Enhancing Tumor), with an overall mean of $\mathbf{8 5. 1 9 \%}$, performing competitively with image-only and text-guided segmentation methods while producing interpretable, semantically grounded predictions.