A Guided-learning Network for Brain Tumor Segmentation with Missing Modalities

Accurate segmentation of brain tumors is essential for clinical assessment and treatment planning, as it leverages multiple MRI modalities to provide complementary diagnostic information. However, missing modalities are frequently en- countered in clinical practice due to factors such as image degradation, artifacts, protocol inconsistencies, or patient con- traindications to contrast agents. While certain unified models are designed to accommodate various missing modality scenarios, their performance declines markedly when multiple modalities are absent, particularly in cases where only a single modality is available. To tackle this challenge, we introduce a Guided- learning Network that leverages dynamic FFT filters. This network trains multiple specialized models designed for different missing modality scenarios, achieving outstanding performance. The architecture comprises two distinct learning paths: one dedicated to capturing multimodal information and another focused on generating modality-specific representations for the missing modality. A guided-learning strategy connects these two paths. Initially, the multimodal path is trained and its parameters are locked, enabling the missing modality path to learn under its guidance. The training process incorporates several key com- ponents: adversarial learning to align high-level features across paths, frequency-domain adaptive filtering using dynamic FFT filters to extract global features, and mutual information transfer to ensure critical information is retained even when modality data is missing. Experiments conducted on the BraTS2018 dataset demonstrate that our model outperforms comparable state-of- the-art methods in all missing modality cases.