[Paper Review] Hierarchical Text-Guided Brain Tumor Segmentation via Sub-Region-Aware Prompts

Bahram Mohammadi, Ta Duc Huy+9 more|arXiv (Cornell University)|Mar 22, 2026

Brain Tumor Detection and Classification0 citations

TL;DR

TextCSP introduces a hierarchical text-guided brain tumor segmentation model with a soft cascade decoder, sub-region-aware prompts, and text-semanthetic channel modulation to improve WT, TC, and ET segmentation on TextBraTS. It surpasses state-of-the-art methods in Dice and HD95 on average.

ABSTRACT

Brain tumor segmentation remains challenging because the three standard sub-regions, i.e., whole tumor (WT), tumor core (TC), and enhancing tumor (ET), often exhibit ambiguous visual boundaries. Integrating radiological description texts with imaging has shown promise. However, most multimodal approaches typically compress a report into a single global text embedding shared across all sub-regions, overlooking their distinct clinical characteristics. We propose TextCSP (text-modulated soft cascade architecture), a hierarchical text-guided framework that builds on the TextBraTS baseline with three novel components: (1) a text-modulated soft cascade decoder that predicts WT->TC->ET in a coarse-to-fine manner consistent with their anatomical containment hierarchy. (2) sub-region-aware prompt tuning, which uses learnable soft prompts with a LoRA-adapted BioBERT encoder to generate specialized text representations tailored for each sub-region; (3) text-semantic channel modulators that convert the aforementioned representations into channel-wise refinement signals, enabling the decoder to emphasize features aligned with clinically described patterns. Experiments on the TextBraTS dataset demonstrate consistent improvements across all sub-regions against state-of-the-art methods by 1.7% and 6% on the main metrics Dice and HD95.

Motivation & Objective

• Motivate leveraging radiological text to improve multi-region brain tumor segmentation.
• Address limitations of single-output-head and global text embedding by exploiting anatomical hierarchies ET ⊆ TC ⊆ WT.
• Develop a parameter-efficient, multi-component framework to align linguistic cues with sub-region segmentation.

Proposed method

• Propose a text-modulated soft cascade decoder with three sequential heads (WT, TC, ET) reflecting anatomical containment.
• Use sub-region-aware prompt tuning with LoRA-adapted BioBERT and per-subregion soft prompts to generate specialized text representations.
• Incorporate text-semanthetic channel modulators (SE-like) to refine decoder features with branch-specific linguistic priors.
• Build on TextBraTS baseline (Swin Transformer visual encoder, BioBERT text encoder, cross-attention fusion, U-Net decoder).
• Train with LoRA on text encoder (query/value projections), small soft prompts (K=4), and text-conditioned SE modules; optimize using SAM with SGD.

Experimental results

Research questions

• RQ1 Can a hierarchical text-guided architecture enforce the ET ⊆ TC ⊆ WT anatomical containment during segmentation?
• RQ2 Do sub-region-aware prompts and LoRA adaptation improve text–image fusion for WT, TC, and ET compared to global text embeddings?
• RQ3 Do text-semanthetic channel modulators provide additional gains by injecting linguistic priors into region-specific feature maps?

Key findings

• TextCSP achieves state-of-the-art average Dice of 87.0% on TextBraTS, outperforming TextBraTS by 1.7%.
• TextCSP achieves best average HD95 of 4.81 mm, improving by ~0.32 mm over the baseline.
• TC gains are largest with TextCSP (+2.6% Dice) among sub-regions.
• Ablation shows full model with soft cascade, sub-region prompts, LoRA, and text modulation yields the highest Dice (87.0%) and lowest HD95 (4.81 mm).
• Sequential WT→TC→ET cascade outperforms parallel or partial cascade strategies in overall Dice (87.0%).