Deformable ConvNets and Muti-Scale Linear Attention for Brain Tumor Segmentation with Missing Modalities

Brain tumor segmentation extracts meaningful information from various modalities (T1, Tlc, T2, and FLAIR), each offering unique insights into brain structure and pathology, essential for accurate tumor subregion delineation. A common clinical challenge is the inability to acquire all modalities due to factors like image corruption or varying protocols. In such cases, most multimodal methods struggle with incomplete inputs, leading to poor performance. This paper proposes a network model, named as DCMLA-Net, using deformable convolution and multi-scale linear attention, effective for any subset of available modalities. Specifically, the deformable convolution extracts modality-specific information, while linear attention fuses multi-scale features, after which the fused features are fed into a deformable convolution decoder to restore spatial resolution and generate the final tumor segmentation. Experiments on the BraTS 2018 dataset show that the proposed method outperforms existing approaches with missing modalities.