A Semi-supervised Teacher-Student Model Based on MMAN for Brain Tissue Segmentation

Salah Felouat, Ramzi Bougandoura, Farrah Debbagh, Mohammed Elmahdi Khennour, Mohammed Lamine Kherfi, Khadra Bouanane

Abstract: In the medical field, deep learning technologies help radiologists gain accurate diagnoses in evaluating both neurological conditions and brain diseases. But these technologies require a huge amount of data to train and validate. The acquisition and labeling of these data are pretty expensive and require medical experts. Therefore, it is necessary to build models that do not require a lot of labeled data like semi-supervised methods. These methods train on few samples of data in a supervised manner and extend this knowledge to the rest of the unlabeled samples. For this reason, we propose a semi-supervised technique called Teacher-Student model using Multi-Modal Aggregation Network (MMAN) and apply it for brain tissue segmentation from Magnetic Resonance Images (MRI). The dataset (MRBrains Challenge) contains a small portion of labeled data and the majority of them are unlabeled. Our proposed method consists of two main steps. The first one aims to exploit the huge unlabeled data to increase the volume of the training set. The second step, it follows the strategy of the Teacher-Student technique in an iterative manner to enhances the performance of our model. The segmentation process will divide the brain into gray matter, white matter and cerebro-spinal fluid. Our approach allows for an improved prediction of brain image segmentation to reach a mean accuracy of 96.21%.