LFC-SATP-SGG: Latent Feature Completion and State-Aware Text Prompting with Semantic-Guided Gating for Incomplete Multi-Modality Brain Tumor Segmentation

Multi-modal Magnetic Resonance Imaging (MRI) is crucial for precise brain tumor segmentation. However, missing modalities are prevalent in clinical practice, severely degrading the performance of deep networks designed for complete multi-modal data. Existing solutions like synthesis or zero-filling often suffer from feature collapse or hallucinated artifacts, while most fusion strategies lack explicit awareness of the specific missing status. To address these challenges, we propose LFC-SATP-SGG, a unified framework comprising Latent Feature Completion (LFC), State-Aware Text Prompting (SATP) and Semantic-Guided Gating (SGG). First, the LFC strategy dynamically generates substitute vectors in the latent space to fill missing dimensions, effectively maintaining feature distribution integrity without the high computational cost of voxel-level synthesis. Second, the SATP mechanism automatically encodes modality availability into semantic text prompts. Utilizing a pre-trained text encoder, SATP provides a semantically structured conditioning signal to guide visual feature interaction via cross-modal attention, eliminating the need for manual input. Finally, SGG is introduced to refine shallow texture representations using deep semantic features for noise suppression. Extensive experiments on the BraTS 2018, BraTS 2020, and BraTS 2021 datasets demonstrate that our method achieves more competitive performance under various missing modality settings. The code is available at https://github.com/MMBTS/LFC-SATP-SGG.