ACN: Adversarial Co-training Network for Brain Tumor Segmentation with Missing Modalities

Accurate segmentation of brain tumors from magnetic resonance imaging (MRI) is clinically relevant in diagnoses, prognoses and surgery treatment, which requires multiple modalities to provide complementary morphological physiopathologic information. However, missing modality commonly occurs due image corruption, artifacts, different acquisition protocols or allergies certain contrast agents clinical practice. Though existing efforts demonstrate the possibility a unified model for all situations, most them perform poorly when more than one missing. In this paper, we propose novel Adversarial Co-training Network (ACN) solve issue, series independent yet related models are trained dedicated each situation with significantly better results. Specifically, ACN adopts co-training network, enables coupled learning process both full supplement other’s domain feature representations, importantly, recover ‘missing’ information absent modalities. Then, two unsupervised modules, i.e., entropy knowledge adversarial modules proposed minimize gap while enhancing prediction reliability encouraging alignment latent respectively. We also adapt modality-mutual transfer retain rich mutual among Extensive experiments on BraTS2018 dataset show that our method outperforms state-of-the-art methods under any situation.