Fig. 2.

Illustration of existing pipelines (a-c) and our proposed framework (d) for segmentation of brain tumors with missing modalities. This figure takes the VS case as an example. Our proposed TISS-Net based on cascaded dual-task networks deals with the synthesis and segmentation tasks end-to-end. A dual-task generator obtains a synthetic target-modality image and a coarse segmentation mask simultaneously. It is followed by a dual-task segmentor obtaining a fine segmentation and errors in the coarse segmentation simultaneously, and a consistency loss between them (i.e, ${\mathcal{L}}_{\mathit{con}}$) is proposed for regularization. Note that the generator is further regularized by a tumor-aware synthesis loss with perceptibility regularization detailed in Section 3.2.